# ocho.bot — Full Documentation ## Roles and permissions Category: Administration Description: The role keys that gate Ocho features, and how subscription tiers work. Access in Ocho is governed by two layers: your company's subscription tier (which applications and features are available at all) and roles (which actions you personally can take). One exception: members of the workspace admin group bypass role checks entirely (see Workspace admins below). Role reference Administrators assign roles on a user's record — the Roles tab on the user's page in the Users table (/table/users). Role definitions themselves, and role assignment where the Users table isn't available, are managed in the admin CMS by a role_admin. Role key · Grants super_rag_dataset_admin — Create and build datasets, upload and delete documents, manage knowledge graphs document_markdown_editor — Create, edit, and publish markdown documents; save chat answers to the Library table_admin — Create custom tables and edit their columns app_admin — Create custom applications in App Builder (the New app button) ui_builder — Use the AI page builder (the New page button and the page-editor wand) citizen_developer — Open the Context Inspector developer tools (the Inspector button in chat and the Context Inspector page) integration_admin — Create, test, sync, reconnect, and delete integration connections auth_admin — Configure sign-in and SSO, user access mode, and allowed email domains branding_admin — Set company name, logo, and primary color during onboarding (the subdomain is shown but cannot be changed after setup) payment_admin — Complete the Billing step during onboarding user_admin — View the Users table and reset user passwords cost_admin — See the company-wide, per-user usage breakdown on the Usage page group_admin — Manage sharing grants to groups (owners of a resource can always share it themselves) role_admin — Manage role definitions and role assignment in the admin CMS flows_author, flows_admin — Create workflows in Ocho Flows (the New workflow action requires one of these) flows_viewer — Reserved for Ocho Flows; assignable but not yet enforced in the app When you lack a role a full page needs, Ocho shows an "Access denied" page that names the required role — contact your administrator to request it. Individual buttons and menu items you lack a role for are usually hidden; where a control is shown disabled instead, its tooltip names the required role. Actions that reach the server may fail with a message naming the required role key (for example "Requires super_rag_dataset_admin role") or, in some cases, a generic "You are not allowed to perform this action." Subscription tiers Tiers order as free < basic < pro < max < enterprise. Navigation items and features carry a minimum tier; anything above your company's tier simply doesn't appear. Workspace admins Members of the admin group bypass role checks entirely: every role-gated page, navigation item, and button renders for them, and server actions accept their requests without the matching role. Admin-group members also get the per-message Inspect context action in chat, on messages where context recording captured the turn. Good to know Some role keys (such as integration_admin and super_rag_dataset_admin) are not part of the seeded role-definition list, so they may need to be created as role definitions before they can be assigned. If the Users table isn't provisioned in your workspace, there is no in-app surface for assigning roles — a role_admin must assign them in the admin CMS. Archiving and restoring tables has no button in Data Manager; those operations are only available through the Data Manager AI chat. They require the table_admin role. In workspaces with row-level security enabled, a table's owner can also archive and restore their own tables; elsewhere, ownership alone is not enough. --- ## App Builder overview Category: App Builder Description: Apps, navigation, and pages — and who can build them. Most of Ocho's sidebar is organized into applications: labeled groups of pages. (A few items sit outside any application — Home at the top, and Get help now, Settings, and your user menu at the bottom.) Alongside the built-in applications, your company can create its own — custom apps with their own navigation and pages, built through chat rather than code. Key concepts Applications — sidebar groups. Which ones appear depends on your company's subscription and each item's role requirements. Pages — what a navigation item opens: a list of records, a record workspace with tabs, or a fully custom page. Custom apps — apps your company owns. They are listed under App Builder > My Apps and served under the private /app/my/ section of your workspace. (The App Builder group also holds SDUI Pages, SDUI Widgets, and SDUI Imports, where the building blocks of custom apps are managed.) The app copilot — each custom app has a chat-driven App Workspace ("Build this app with me") that can attach data tables and pages and shape the navigation your teammates see. Changes always ask for your approval first. The UI Builder — a separate chat-driven page copilot that generates and modifies an app's custom pages; you describe what you want, review the draft, and publish it live. Who can do what Creating apps requires the App Admin role. Building pages requires the UI Builder role. (A new app's Create your first page button additionally requires the UI-builder feature on your subscription.) Good to know The builder controls only appear if you hold the matching role — being a workspace admin is not enough on its own. If you don't see New App or the page-builder controls, ask an admin to grant you the App Admin or UI Builder role. A brand-new custom app opens to a placeholder page titled "Your new app" until a page is built for it. Users without the UI Builder role see: "Ask an admin to grant you the role, or contact 2men for whiteglove implementation." Continue with Build custom apps. --- ## Flows overview Category: Ocho Flows Description: Visual, multi-step automations over your documents, data, and AI. Ocho Flows lets you build multi-step automations on a visual canvas: fetch documents, run them through the AI, call external integrations through your configured connections, loop over items, and assemble a report — then run it manually or let a trigger start it for you. As the Workflows list puts it: "Visual DAG workflows triggered by manual run, webhooks, schedules, or record events." Key concepts Workflows — named automations, each edited on a drag-and-drop canvas of connected blocks. See Build workflows. Blocks — the steps: AI calls (LLM Call, Chat Skill), dataset search, document readers, integration calls, human approval gates, loops, and more. There is no generic HTTP, code, or condition block. See the block reference. Runs — each execution of a workflow, with a live event timeline and a markdown run report. See Run and monitor. Triggers — how a workflow starts: manual run, a webhook, a schedule, or a record event (a change to a Data Manager record or a Library document). Only record-event triggers can be edited in the editor's Triggers dialog; triggers arm on publish. See Triggers and connections. Connections — authenticated links to external services that Integration Call blocks use via the connector registry. Flow Copilot — an AI assistant panel in the editor that builds and modifies workflows from plain-language descriptions, where it's available in your workspace. Three tabs The Flows section has Workflows (create and open them), Runs ("Workflow execution history and status."), and Connections ("Manage integration connections used by your flows."). Flows and chat The main Chat assistant may propose a workflow when your request is multi-step. When it does, the proposal appears as a card in the conversation asking "Run this multi-step workflow?" — it never runs by itself. Use the card's View, Edit on Canvas, or Run buttons to review, edit, or run it explicitly. Depending on the chat mode, the assistant may instead answer with a plain-text workflow design rather than a runnable proposal. Who can do this Creating workflows is permission-gated. If your account lacks the flow permission, Create in the New workflow dialog fails with "Could not create workflow: You are not allowed to perform this action." Ask a workspace admin to grant you access. Good to know Workflows currently can't be deleted or duplicated from the interface. The canvas has no undo — changes (including Copilot edits and node deletions) apply immediately and auto-save. The toolbar also has a manual Save button. Publish checkpoints a version of your workflow — and it's also what arms triggers: "Triggers arm on publish." Older workflows that predate the canonical format can't be edited: opening one shows "This workflow predates the canonical format — its old canvas can't be edited. Start fresh or recreate via chat." with a Start fresh button. --- ## Data Manager overview Category: Data Manager Description: Build your own structured tables and let the AI work with them. Data Manager is a workspace for custom tables — structured data your team defines, like a lightweight database with an AI assistant attached. It has three resizable panels: a Tables list on the left, the selected table's records in the middle (until you pick a table it reads "Select a table on the left."), and a Data chat pane on the right with the prompt "Ask about your data...". Key concepts Tables — created by table administrators, each with a label, icon, and description. Columns are added afterwards from the table's Manage columns option (the Columns sheet, where admins add, edit, reorder, and delete columns). Columns — typed fields. The available types are: String, Text (long), Number, Boolean, Date, Date + time, Choice (single), Choice (multi), Reference (link to another table), Currency, URL, Email, Phone, and JSON. AI access — when creating a table, the administrator sets two checkboxes (both off by default): Let AI read this table (chat skills can list and summarize records) and Let AI write this table (chat skills can create and update records). With write access, the assistant creates and updates records for you directly — there is no separate approval step, so grant write access deliberately. Who can do what Table administrators (the table_admin role) create tables and manage their columns. Archiving and restoring a table is done by asking the AI assistant — there are no archive or restore buttons in the workspace, and archiving is reversible. Everyone with access to Data Manager can use the Data chat. Which tables you can see, and whether you can add or edit records, both depend on that table's permissions and visibility, set per table by an administrator — a restricted table is visible only to its owner, table admins, and shared groups. Good to know The Data chat pane is the standard Ocho chat surface pointed at your data — its empty-state suggestions and dataset picker are the same ones you see in regular chat. If your workspace has no tables yet, the page shows "No tables yet." Table administrators also see a "Create your first table" button; everyone else sees "An admin can define one in Payload > Data Manager." — ask a table administrator to create the first table. The pages Create custom tables Manage records The AI data assistant --- ## Browse, filter, and sort Category: Working with tables Description: Work any list page — search, filters, columns, and saved views. The Documents, Datasets, and Connections list pages share the same table toolkit: a search box, Filter, a row-density button, Columns, views, and Rows per page. Individual tables add their own actions on top — for example, Datasets has a Create dataset button, and tables that allow record creation show a create button. Search and filter The search box matches the Name column only — it won't find text in other columns like Source or State. Its placeholder names the table, e.g. "Search Documents…". Filter opens a builder ("Filter rows"). Click + Add condition, then pick a column, an operator, and a value. The operators fit the column type: Text columns: contains, does not contain, is, is not, starts with, is empty, is not empty. Date columns: is, >, ≥, <, ≤, between, is empty, is not empty, with a date picker for the value. Choice columns (like State) switch the value input to a Select… dropdown of the allowed values. Add as many conditions as you like — the Filter button shows a count badge (e.g. "Filter 2") while conditions are active. Conditions on different columns apply together (AND). Active filters show as removable chips under the toolbar, with a Clear filters link. Sort and shape the table Click a column header to sort — ascending, then descending, then off. One column sorts at a time. The Columns button opens a menu where you can show/hide each column with its checkbox, reorder with the drag handle or the up/down arrows, and re-enable everything with Show all. The row-density button (icon only, next to Columns on the right side of the toolbar) switches between Compact, Comfortable, and Spacious rows. Rows per page offers 10, 25, 50, or 100 (default 25), with Previous/Next page controls and a range readout like "1–25 of 519". Saved views The views button in the toolbar reads Default view until you save and apply a view of your own. Saving a view captures your search, filters, sort, visible columns, density, and page size as a named entry under SAVED VIEWS in that menu. Click a saved view's name to apply it; the active view shows a checkmark and its name appears on the toolbar button. Change any setting while a view is active and a small dot appears on the button; choose Update "your view name" in the menu to save the changes into the view. Hover a saved view to reveal Set as default, Rename, and Delete. Delete removes the view immediately — there's no confirmation. Saved views currently last only for your browser session. After a full page reload the SAVED VIEWS list is empty again, and a view marked as default won't be applied on your next visit. For anything you want to keep or share, use the URL instead (below). Deep-link everything Your search, filters, sort, page size, and page number live in the URL — copy it to share the exact slice of data you're looking at. Opening a copied URL restores the search text, filter chips, sort, and pagination exactly. Good to know Filters combine with AND only; there's no "or" between conditions. Two conditions on the same column don't combine: the newest one takes effect and the earlier one is ignored, even though both chips stay visible. Use different columns per condition. When a search or filter matches nothing, the table shows the generic "No rows yet" empty state with a "No results" counter — it looks like an empty table. Clear the search box or click Clear filters to get your rows back. Some pages are read-only or restricted. A restricted table shows an "Access denied" message naming the role required — contact your administrator to request access. A read-only record says so and points you to where it can be edited. If a table fails to load you'll see "Couldn't load schema for …" — that's an error, not a permissions message; let your administrator know. --- ## Library overview Category: Documents & datasets Description: How documents and datasets work, and who can manage them. The Library is where your company's knowledge lives. In the sidebar it is the LIBRARY group, with two items: Documents — the individual documents Ocho indexes. The Source column shows where each one came from: Manual Upload (files you upload), Integration (synced from a connected tool such as Notion), or In App (markdown documents written in Ocho). Ocho parses, chunks, and indexes each one. Datasets — named collections of documents that chat and search query. A document can belong to many datasets; attaching never copies or moves it — upload once, attach to any dataset. Document lifecycle Documents move through processing stages — Parsing, Extracting, Chunking, Embedding, Indexing — shown as per-stage progress while a document is in flight. Once processing finishes, the document's Lifecycle section shows just its state. States: State · Meaning Pending / Processing — Working through the pipeline Draft — An in-app markdown document that hasn't been published yet Indexed — Ready for datasets, chat, and search Failed — Something went wrong; the document shows the error and a suggested fix (for example "Retry now" or "Enable Force OCR and rebuild") Dataset lifecycle Datasets are built, not just filled. A journey stepper on each dataset walks you through: Add documents → Build → Review → Use in chat. States along the way: State · Meaning Draft — New; attach documents Scanning / Proposed / Approved / Building — Ocho is analyzing content and constructing the search index Live — Ready — chat and search can use it Archived — Taken out of service; documents remain in the Library Failed / Cancelled — The build stopped; you can retry Who can do what Viewing documents, datasets, and knowledge graphs is open to everyone in your workspace. Changing things is role-gated, and the server enforces it — for example, an upload without the right role fails with "Upload failed: Requires super_rag_dataset_admin role": Dataset administrators create and build datasets, upload and delete documents, and manage knowledge graphs. Markdown editors create and publish markdown documents. See Roles and permissions. Good to know Only Live datasets appear in the Search page's dataset map and the chat Sources picker; draft datasets stay out of both until built. The Documents list doesn't show which datasets a document belongs to. To see what's attached, open the dataset and use its Documents tab ("Attached documents"). Search-quality evaluations still run automatically after each dataset build, but there is currently no way to view or manage them in the app — the dataset's QA tab does not load a view. The pages Upload documents Create and build datasets Manage datasets Knowledge graphs Markdown documents Evaluate search quality --- ## Search overview Category: Search & knowledge graph Description: Direct document search paired with an interactive knowledge graph. The Search page finds passages across your datasets and shows how your knowledge connects. It pairs a ranked results list with an interactive knowledge graph of the entities and relationships Ocho extracted from your documents. What you see Before you search — the page shows the dataset map: one cell per dataset, sized by document count, with the dataset name and a green check on each selected cell. Deselecting datasets in the picker dims their cells rather than hiding them. Clicking a cell dives into that dataset's full knowledge graph — but only for datasets that have a graph on Search; "no graph" cells don't respond to clicks. A Back to map button (top left) returns you to the map. Searching — search runs automatically as you type into the Search across selected datasets… box; there is no Search button. Typing a couple of characters also opens a suggestions popover with Entities (with mention counts) and Documents — clicking an entity suggestion opens its detail view, and clicking a document suggestion opens that document in a new tab. (Pressing Enter doesn't pick a suggestion — click the one you want.) While you search — the page splits into a graph pane and a results pane (headed RESULTS). The graph automatically drills into the top result's dataset and focuses that document's node. Hover a result card to light up its document node in the graph. You can resize the panes with the 30/70 split, 50/50 split, and 70/30 split buttons at top right; your choice is remembered. Result cards offer two actions: Open source and Chat about this (clicking elsewhere on the card does nothing). Open source opens the same source viewer used by chat citations. For PDF sources the matched passage is highlighted on the rendered page; for markdown sources you get a one-line excerpt summary above the document instead of an in-document highlight. A mini-graph sits in the left sidebar above two tabs: Graph neighbors and Search results — the latter is an editable mini-search prefilled with your query, listing related passages across datasets. When is a dataset searchable? Search needs a dataset that is fully built — draft datasets don't appear in the picker at all. A knowledge graph is not required to search: datasets show a small "no graph" tag in the picker, but they are still selectable and return results. What the tag reliably means is that the dataset's map cell can't be clicked into a graph (its tooltip reads "No knowledge graph yet"); after a search, the page may still auto-drill into such a dataset and show whatever graph data exists for it. (A dataset can carry the tag here even though its Overview tab shows a built graph.) Graphs are generated manually by a dataset admin from the Knowledge Graph card on the dataset's Overview tab; see Knowledge graphs. Good to know The dataset picker button reads All datasets (or N of M datasets) and includes a Filter datasets… box; once you deselect a dataset, a N of M selected header row with a Select all link appears. You can't deselect the last remaining dataset, and the picker is hidden while you're drilled into a single dataset's graph (use Back to map first). Your search, dataset selection, and focused entity are reflected in the page URL — the browser Back button retraces your steps, and you can bookmark or share a search. The query and dataset selection are restored from a shared URL, but the focused entity is not: the search re-runs and refocuses the top result. Search returns up to 20 passages per query. On phones and small screens the graph moves behind a View graph button; the results list takes the full width. --- ## Chat overview Category: Chat Description: How conversations, sources, and citations fit together. Chat is a conversational AI assistant grounded in your company's documents. Ask a question and Ocho searches the datasets you've selected, composes an answer, and cites the exact documents — and pages — the answer draws on. Key concepts Conversations — every chat is saved to your private history, listed in the sidebar by date. You can pin, rename, delete, and search them. See Start and manage conversations. Sources — the datasets the assistant searches. You control them from the Sources button in the composer; your selection applies to the messages you send next (in any conversation) and resets to all sources when you reload. See Choose data sources. Citations — each grounded answer shows a sources row underneath (for example, 9 sources). Open it for the Citations from answer panel, where numbered citations link to the exact source passage in an in-app document viewer. See Citations and sources. Modes — a conversation can run under different instruction prompts ("modes"), and you can tune temperature, response length, and model per message. See Tune responses. Message actions — copy an answer, or copy a link to it once the conversation is saved. Users with the document_markdown_editor role also get a Save to Library action. See Message actions. While Ocho responds A status line shows what's happening: Searching documents, the internal id of a tool being run (for example, search_knowledge), then Thinking — replaced by the answer as it streams in. The Send button becomes a Stop button during a response. Stopping halts the streaming in your current view and temporarily marks the reply "stopped" — but the full answer is usually generated in the background anyway, and that complete answer is what you'll see when you return to the conversation. Chat doesn't yank the page down while text streams. When you scroll away, a floating down-arrow button (labeled "Jump to latest" for screen readers) appears above the composer — click it to jump to the newest content. Good to know Answers can be wrong. As the composer notes: Ocho is AI and can make mistakes — double-check responses, and use the citations to verify. If no relevant passages are found, Ocho says so rather than inventing a source. If document search fails, the answer is generated without documents and flagged with a warning. Very long answers can hit a length limit — a notice suggests asking a follow-up to continue. File attachments in chat are not yet available. The composer shows a paperclip Attach file button, but clicking it only shows "File attachments coming soon". Chat has no search-only mode. To search documents without an AI answer, use the separate Search page in the sidebar. --- ## How Ocho compares Category: Market comparison Description: Where Ocho fits among second brains, note tools, and AI research assistants. Every tool in this space asks you to do the same thing: start over. Write your notes in our app. Upload your files to our notebook. Move your docs into our workspace. Ocho starts from the opposite premise — your company's knowledge already exists. It's in SharePoint, Notion, Google Drive, GitHub, email threads, and file shares. The problem isn't creating a second brain; it's that the first one is scattered, unsearchable, and walking out the door with every departing employee. Ocho connects to what you already have, ingests it continuously, builds the datasets and knowledge graph automatically, and gives your whole team one AI experience over all of it — with citations you can check. Start here Why do I need Ocho? — the case, with numbers. The concepts What is a knowledge graph? What is agentic ingestion? What is agentic RAG search? The comparisons Ocho vs. Karpathy's second brain Ocho vs. Obsidian Ocho vs. NotebookLM Ocho vs. Zaro AI Ocho vs. Notion Each comparison is honest about what the other tool does well. They're good tools. They're just built for a different job. --- ## What is Ocho? Category: Getting started Description: A quick tour of the Ocho platform and its core concepts. Ocho is an AI workspace for your company's knowledge and data. You upload documents, Ocho indexes them into searchable datasets, and from there you can chat with an AI assistant that cites its sources, explore an automatically built knowledge graph, manage structured data in tables, and automate multi-step work with visual workflows. Core concepts Workspace — your company's private Ocho environment, reached at your own web address. Everything in it is scoped to your organization. Applications — the labeled sections in the left sidebar. Built-in sections include OCHO (with Chat and Search), LIBRARY (with Documents and Datasets), OCHO-FLOWS (with Flows), DATA MANAGER, APP BUILDER, and INTEGRATIONS; any custom apps your team builds appear as their own sections alongside them. Which applications you see depends on your company's subscription and your role. Documents — files you upload (.pdf, .docx, .pptx, .xlsx, .xls, .csv, .txt, .md, .html, and .vtt/.srt transcripts). Ocho parses and indexes them so the AI can read them. Datasets — named collections of documents. A dataset must be built before chat and search can use it. One document can belong to many datasets. Chat — a conversational assistant that answers questions using your datasets and cites exactly where each answer came from. Search & knowledge graph — direct search across your documents, paired with an interactive graph of the entities and relationships Ocho extracted. Flows — visual, multi-step workflows (call the AI, search datasets, call APIs, loop over documents) that run on demand, on a schedule, or on events. Roles — fine-grained permissions. Many actions (creating datasets, building workflows, administering users) require a specific role. See Roles and permissions. Where to start 1. Sign in to your workspace. 2. Find your way around the sidebar, settings, and help. 3. Upload a few documents and create a dataset. This step requires the super_rag_dataset_admin role — if you don't have it, ask an administrator to create and build a dataset for you. 4. Start a conversation and ask a question about them. Good to know Uploads accept files up to 5.00 GB per file, with no file-count limit. Legacy .doc and .ppt files are not accepted — convert them to .docx/.pptx first. The Chat surface can be branded per workspace: its title and input placeholder may be customized for your company (for example, a chat titled after your document collection), but it works the same everywhere. Chat's Sources popover also lets you tune each conversation — temperature, response length, model, and mode. See Tune responses. Role checks are enforced on the server: creating datasets or workflows without the required role fails with an error such as "Requires super_rag_dataset_admin role", even for workspace administrators who haven't been granted that role. --- ## Sign-in and SSO Category: Administration Description: Configure how your organization signs in — passwords, SSO, and who may join. Who can do this: auth administrators. The Auth Management card on Home opens the page that controls your organization's sign-in. The page itself is headed Authentication — "Manage how your organization signs in and how new users get access." See the current method The top card, Current method, shows what's active: Email & Password, or Single Sign-On (OIDC) / Single Sign-On (SAML) with the provider details. Control who can join Two access modes: Invite only — only people you add can sign in. Self-service — anyone from an allowed email domain is provisioned on first SSO login. When selected, manage the Allowed email domains list right below — self-service is blocked until you add at least one domain. Changes here save immediately: picking a mode or adding/removing a domain takes effect on click — there is no Save button. Set up SSO 1. Copy the displayed redirect URI and register it with your identity provider. 2. Choose OIDC (client ID, client secret, issuer URL) or SAML (metadata URL, or entity ID + SSO URL + certificate). 3. Optionally click Test connection to validate the configuration without a live login. 4. Click Enable SSO. If SSO is already configured, the card and its button read Update SSO instead. Switch back to passwords Switch back to email & password turns SSO off. Because SSO users have no password yet, the confirmation dialog offers Email new passwords to affected users now (checked by default); confirm with Switch to basic auth. Existing sessions remain valid until they expire. Good to know Secrets are write-only — the page never redisplays a saved client secret; updating SSO means re-entering the credentials. During company onboarding, the same setup is offered as the Security step, including Delegate to Someone — enter the colleague's full name and email (both are required). --- ## Build custom apps Category: App Builder Description: Create an app and shape its pages with the AI page builder. Create an app Who can do this: users with the App Admin role. Without it, the New App button is hidden entirely. 1. Open App Builder → My Apps in the sidebar and click New App. 2. In the New App dialog, enter a Name (the app's identifier is derived from it automatically), an optional Description and Icon, and set AI access with Let AI read this app's data and Let AI write to this app's data. 3. On create, you land in the App Workspace: the app's name with an Active badge and an Open app button in the header, a Build: (app name) copilot chat on the left, and a tab group on the right that switches between Navigation, Data tables, and Pages lists. Describe what you want the app to do — the copilot can add navigation, attach data tables, and wire up pages. 4. Click Open app to view the app itself. If you can build pages, its empty overview shows a Create your first page call-to-action; otherwise it explains that you need the UI Builder role. Build pages with AI Who can do this: users with the UI Builder role. App Admins can also publish page changes via the API. 1. On a custom page, open the builder with the wand button in the header (Edit this page with AI). The wand appears only on custom pages, and only for builders. 2. The UI Builder panel docks to the right with a Page copilot chat. Describe the page you want or the change you'd like to make. 3. The AI edits the page's draft; the page re-renders as changes land. 4. Click Publish draft to live to make the draft live for everyone. You can also create a page without going through an app's overview: the App Builder → SDUI Pages list has a New Page button (role-gated) that opens the New SDUI Page dialog — pick the app, name the page, and click Create. Drafts, versions, and safety nets Edits accumulate in a draft that only builders and App Admins see until published. The publish button's label reflects the page's publish status — Publish draft to live if the page has never been published, Published (re-publish to sync draft) once it has (even if the draft has newer edits) — and each page's detail view shows draft/published and build status badges. The Versions tab on a page's detail view shows the last 20 versions in its Version history list. Every prior version has a Restore button; the current one is marked LATEST. If a page build fails, viewers keep seeing the last good version while the builder shows a Build failed error; ask the AI to fix it, then publish again. Good to know Publishing is blocked while a build is in progress or failing. Custom pages compose from Ocho's widget catalog — layout containers, cards, buttons, chat panels, document viewers, data tables, and more — so generated pages match the rest of the product. You can browse the full read-only catalog under App Builder → SDUI Widgets. --- ## Build workflows Category: Ocho Flows Description: Compose automations on the canvas — by hand or with the Flow Copilot. Who can do this Creating workflows can be restricted server-side. If clicking Create shows "Could not create workflow: You are not allowed to perform this action.", ask an administrator to grant you access — admin group membership alone does not guarantee it. Create a workflow 1. On the Workflows tab, click New workflow. Give it a Display name and a Slug (lowercase letters, numbers, dashes, and underscores; 2–63 characters; must start with a letter). The display name is not filled into the slug automatically, and an invalid slug simply keeps the Create button disabled rather than showing an error. 2. The canvas editor opens with an empty canvas — drag in your first block from the palette to get started. Work the canvas Add blocks by dragging them from the left Blocks palette onto the canvas. Connect blocks by dragging from one block's output handle to another's input. Each block also shows a second, red Error output handle, but it doesn't route anything today — when a block fails, everything downstream is skipped, including anything wired from the Error handle. Configure a block by clicking it — an inspector opens with its name and settings (prompts, queries, model choices, document and dataset pickers, and so on). Delete a selected block or connection with Backspace/Delete. A minimap and zoom controls help on large flows. Everything auto-saves a couple of seconds after you stop editing; a Save button is there when you want to be explicit. Reference earlier results Block settings can reference the output of any earlier block with double-brace expressions — for example {{ search.results }} or {{ vars.total }}. The block reference lists the available patterns. Use the Flow Copilot The Flow Copilot panel is already open on the right when the editor loads — an AI that sees your current canvas. The Assistant and Run buttons in the toolbar switch what the panel shows; the panel itself can't be closed. With Assistant selected you can: Describe a flow from scratch — "for each document in dataset X, summarize it and compile a report" — and the Copilot lays out the blocks and wiring. Ask for changes to what's there — "add an approval step before the report". Copilot changes apply to the canvas immediately. Publish a version Publish opens the Publish release dialog, where you set a Version label and an optional commit message, then snapshots the workflow. Publishing the same content twice doesn't create duplicate versions — you'll see "No changes since last commit; nothing published." Since the canvas has no undo, publish before risky edits. Good to know Run saves any unsaved changes and starts the run immediately — there is no pre-run check or warning. Blocks that can't run are marked skipped during the run; the run finishes with a partial-failure status if at least one block succeeded, and as failed if nothing succeeded. A workflow's slug can't be changed after creation. A few advanced block settings are edited as raw JSON for now. --- ## Create custom tables Category: Data Manager Description: Define tables and columns for your team's structured data. Who can do this: table administrators. Without the Table Admin role these controls simply don't appear — the left rail shows only the Tables heading, and with no tables the empty state reads "No tables yet." with a note that an admin can define one. Create a table 1. In Data Manager, click New next to the Tables heading in the left rail (with no tables yet, an additional Create your first table button also appears in the empty state). This opens the New table dialog. 2. Give it a Label (the internal name is derived automatically as a snake_case slug — there is no name field), an optional Description, and an Icon. 3. Decide AI access: Let AI read this table and/or Let AI write this table. Both are off by default. The read toggle controls whether the AI data assistant can read the table, and the write toggle controls whether it can create and update records. Deleting records through the assistant is governed by your own delete permission, not by these toggles. 4. Click Create table. Define columns Open the Columns sheet with the Manage columns gear button on the table (or the Add column button, which opens the same sheet with the add form first): Add a column — set a Label, a Name (snake_case, derived from the label but editable here), a Type, plus Mandatory and Visible to AI checkboxes (Visible to AI is on by default). Choice columns — Choice (single) or Choice (multi) — take a comma-separated list of choices; Reference columns pick a Reference table; String and Text (long) columns can set a Max length. Reorder columns with the Move up / Move down controls. Edit a column at any time — label, type, max length, choices, reference table, mandatory, and visible to AI can all change after creation. Delete a column — it is removed from the schema; existing record values are retained in storage but no longer visible. Good to know A column's internal name can only be set when the column is added — the edit form never offers a rename, so plan names when you create columns. The label stays editable at all times. Deciding AI access happens at table creation; there is no in-app control to change it afterwards, so choose deliberately for tables holding sensitive data. --- ## Edit records Category: Working with tables Description: Create and edit records, link related records, and keep notes. Who can do this Most tables can be browsed and read by everyone, but a few (such as Users) require a role just to view. Creating, editing, deleting, and linking records requires write access to that specific table, granted by a per-table role (for example user_admin, app_admin, or super_rag_dataset_admin). Without it, a table is read-only: no New record button, no bulk actions, and no editable form. Being a workspace admin does not by itself grant table write access. Create and edit On tables where you have write access: New record in a list's toolbar opens a form generated from the table's fields; required fields are marked with a red asterisk. (Some tables replace or hide this button with their own action — for example, Datasets has a Create dataset button and Documents has an upload dropzone instead.) Editable record pages show the form; read-only fields display below it. Select rows in a list for bulk actions: Delete (confirmation reads "This will permanently remove {n} rows. This action cannot be undone.") or Update, which lets you pick a field and a new value to apply to every selected row. Clicking a row doesn't always open an edit form. Many tables — Connections, Documents, Datasets — open an overview page instead, with metadata tiles and related-record tabs. And on tables you can't edit, the record page shows "This record is read-only here. Edit it in the Payload admin." above the raw record data. Related records Record overview pages can show related-record tabs — for example, a connection's Sync History and Synced Documents tabs, or a document's Attached Datasets tab. Each tab is a full mini-table with its own search, filters, and paging, plus (when you have write access to the related table): Add {table name} — either creates a related record inline, or opens a picker of existing records to link (for many-to-many relationships). Select several in the picker and use Add selected to link them in one go. Click a row in a one-to-many list to edit it in a dialog without leaving the page. Linked (many-to-many) rows can't be edited from the tab — clicking them does nothing. Select linked rows and use Delete to remove the links. The confirmation uses the same permanent-delete wording as bulk delete, but on a link tab it removes only the link, not the underlying record. Record notes Where enabled, records carry a Notes journal: Write notes in the rich-text composer ("Add a note…") — markdown like bold and lists renders as you type. Add note posts instantly. Only the author can edit or delete their note; edits are marked "(edited)", and hovering shows when. Deleting a note removes it from the list, but its history is preserved. Search notes ("Search notes…"), filter by author, and sort by **Last updated (the default), Oldest updated, Newest first, or Oldest first. Long notes collapse behind Show more, and lists of more than 25 notes paginate behind a Load more** button. Good to know Notes unlock after the record is first saved — a new, unsaved record shows a lock with "Save this record to enable notes" until then. The link-existing-records picker loads up to 200 candidates, and its search box only filters those 200 — on very large tables, records beyond that cap won't appear. Deleting records from a bulk selection cannot be undone. On read-only tables, selecting rows still shows the "{n} selected" bar, but with no actions — that's expected, not a bug. --- ## Upload documents Category: Documents & datasets Description: Add files to the Library and track their processing. Who can do this: Library admins. The upload zone is visible to everyone, but uploads from anyone else fail with "Upload failed: Requires super_rag_dataset_admin role" after clicking Upload. Upload files 1. In the sidebar, open Documents under Library. 2. Drag files onto the upload zone ("Drag documents here or click to select"), or click it to pick files. 3. Picked files are queued, not uploaded yet — each shows its size and a Queued badge. Use Remove to drop a single file or Clear to empty the queue. 4. Click Upload (N) to start the batch. Each file shows a spinner with a percentage while its bytes upload, then moves through Parsing → Extracting → Chunking → Embedding → Indexing until it lands on Indexed (or Failed). Supported types: PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx, .xls), CSV, plain text (.txt), markdown (.md), HTML, and transcript files (.vtt, .srt). Size limits: up to 5 GB per file. There is no file-count limit, but a batch can't exceed 5 TB in total — files beyond that are skipped with "Batch total can't exceed 5.00 TB; N file(s) skipped." Track and fix documents The Documents table shows every file with its State — Queued, Parsing, Chunking, Embedding, Indexing, Indexed, or Failed. (The Extracting stage from the upload zone is folded into Parsing here, and in-app markdown documents that haven't been published show a draft state.) While a batch is indexing, Library admins see an "Indexing in progress" banner above the table with a completed/total counter and progress bar, and a matching "Indexing N/M" pill in the footer tracks the batch from any page. Keep the tab open until each file's byte upload finishes; after that, parsing and indexing run server-side and you can safely navigate away. On a document's page, the Lifecycle section shows where it is. While a document is processing, it displays a stage-by-stage stepper with the current stage highlighted. An estimated finish ("~Ns remaining") appears only when timings from earlier runs are available — documents processing for the first time typically show no estimate. Once finished, it shows just the state chip (for example Indexed). For failed documents, the page shows the error with a suggested fix — for example enabling Force OCR for scanned files, or simply retrying. As a Library admin, clicking a fix button (for example Enable Force OCR and rebuild or Retry now) starts the rebuild immediately — no Reason field needed. For everyone else the buttons appear disabled, and on the org-wide document library's document page they are not wired up at all. Rebuild Document — Library admins can re-run processing from the Admin Actions section on the document page, optionally ticking Force OCR (re-extract text via OCR even if the PDF has embedded text). You must fill in the Reason (required) field first — the Rebuild button stays disabled until you do. Add context to a document Editing document context currently lives in the org-wide document library, which is not in the sidebar yet — open /app/library/documents directly and click through to a document's overview. Its About section is collapsed by default (it shows just "No description yet." until you click the About header to expand it). Once expanded, it has an editable description ("Add a short summary so this is easy to find later." — saves on blur) and rich-text Notes (bold, italic, bullet and ordered lists, with a Save notes button). The document page reached from the sidebar Documents table shows the description read-only. Delete documents Deleting removes a document from every dataset it belongs to and deletes its chunks and embeddings. This cannot be undone. Delete is a Library-admin action, and it currently lives only in the org-wide document library (/app/library/documents): open a document's overview to delete it — the confirmation there shows how many datasets are affected ("This document is attached to N dataset(s)…") — or select rows in that list for bulk delete, which asks only a generic "Delete N rows?" without a dataset count. The sidebar Documents table has no delete — its only selection action is Add to dataset. Good to know Uploading a document does not make it searchable by itself — it must belong to a dataset that has been built. See Create and build datasets. Individual failed files in a batch can't be retried in place yet — the Retry button is disabled ("Retry will be added in a follow-up"). Remove the file and re-upload it. The Attached Datasets list on a document's page currently shows "No rows yet" even for documents that are in a dataset. To see what's attached, open the dataset and check its Documents tab instead. The "Search documents…" box in the org-wide document library does not currently filter the list. When you re-publish or reprocess an indexed document, the previous version keeps serving search results until the replacement finishes indexing — the document keeps its Indexed badge while a smaller re-indexing progress bar runs. --- ## Search your documents Category: Search & knowledge graph Description: Query datasets, scope by dataset, and work the results list. Run a search 1. Open the Search page. The resting view is the dataset map — a full-screen graph of your datasets. 2. Scope the search with the dataset dropdown to the right of the search box. It reads All datasets (the default) or N of M datasets. Clicking it opens a list with a checkbox per dataset, each showing its document count. Type in the Filter datasets… box to narrow a long list by name, and use Select all to restore everything. At least one dataset must stay selected — the last checked row can't be unchecked. 3. Type your query. The search runs automatically a moment after you stop typing (no submit button). Running a search drills the graph into the top hit's dataset and shows the results next to it. While drilled in, the dataset dropdown is hidden — click Back to map to change your dataset selection (clearing the query closes the results but keeps you drilled in). Reloading doesn't help: the search re-runs and drills right back in. Typeahead From two characters on, a dropdown suggests: Entities — people, materials, concepts extracted into the graph, with their type and how often they're mentioned (for example "PROCESS · 6 MENTIONS"). Clicking one opens its top passage in the source viewer, with Graph neighbors and Search results tabs alongside. For PDF sources the passage is highlighted in the full document; other formats (like markdown) show the full document, sometimes with a one-line gist banner when a summary is available. Documents — matched by exact filename text; selecting one opens the document in a new tab. Note the match is literal, so "100m offers" won't find "100m_offers.pdf". Suggestions are selected by clicking only — pressing Enter does nothing, even with the dropdown open. The search itself runs automatically anyway. Read the results Each result row shows a color-coded relevance percentage, the source file and page (like "100m_offers.pdf · p80"), and a snippet. Hovering a row highlights that document's node in the graph. Two actions per row: Open source — view the passage in the full document. PDF sources show it highlighted; other formats open without a highlight (a one-line gist banner appears when a summary is available). Chat about this — a "Chat about this source" dialog asks whether to start a New chat with the excerpt as the prompt or add it to an Existing chat (your last 10 conversations). The chat composer arrives prefilled with the quoted passage, ready for your question. Adjust the layout On desktop, three preset buttons set the graph-to-results split (30/70, 50/50, 70/30). Your choice is remembered across sessions and devices. Good to know Some datasets show a small "no graph" tag in the dataset dropdown. They stay fully selectable and searchable, and their graph can still render when a search drills into them — the tag only stops you from clicking their cell on the dataset map to drill in manually. Results are capped at the 20 best passages — refine the query rather than paging. Search always returns the closest passages it can find, so a query that's off target shows up as low relevance percentages rather than an empty list. If everything scores low, check that the right datasets are selected before rewording the query. --- ## Start and manage conversations Category: Chat Description: Send messages, control streaming, and organize your chat history. Start a conversation 1. Open Chat from the sidebar. 2. Type your question in the composer. Enter sends; Shift+Enter adds a new line. On a fresh chat you can also click one of the suggested question tiles. 3. Watch the answer stream in. While a response is streaming, the send button turns into a red stop button (a white square icon) — click it to cut the response short. Stopping only halts streaming in your current view: the partial text and its stopped badge are shown there but are not saved. When you come back to the conversation, the turn shows the full answer the server finished generating — or, if you stopped before any text arrived, the turn may not appear at all. Your conversation history The conversation list (left of the chat on desktop, inside the sidebar drawer on mobile) groups chats into Pinned, Today, Yesterday, Previous 7 days, and Older. From each conversation's row you can: Pin / unpin — pinned chats stay at the top. Rename — edit the title inline; press Enter to save. Delete — removes the conversation after a "Delete this conversation?" confirmation. Use the filter box (placeholder "Search chats…") to find a conversation by title. Conversations are titled automatically: shortly after the first answer, the title in the list is replaced by a generated one based on your question. If you rename a chat right after its first message, the automatic title may overwrite yours — rename again once it settles. Click New chat at the top of the list to start fresh at any time. The conversation title in the chat header is also a dropdown with its own New chat option. Search inside a conversation Once a conversation has messages, a search icon appears in the header. It opens a find bar that matches text across all messages. Matching works at the message level: the counter shows how many messages contain your text, and each step scrolls the matching message into view and flashes it (individual occurrences inside a message are not highlighted). Step through matches with the arrows or Enter / Shift+Enter, and close with Escape (with the find bar's input focused). Good to know Your current conversation is restored automatically when you come back to the page. Each answer (assistant reply) has a stable link — see Message actions — so you can bookmark or share a specific answer with a teammate. Your own messages don't have links or actions. Deleting a conversation cannot be undone. --- ## Why do I need Ocho? Category: Market comparison Description: The cost of scattered knowledge, and what changes when your team shares one brain. Your company already paid to create its knowledge — every proposal, spec, contract, transcript, and post-mortem. Then it paid again to lose it. The problem, in numbers Knowledge workers spend nearly 20% of the workweek — a day a week — hunting for internal information or the colleague who has it (McKinsey). A 2025 survey puts it at 3 hours a day across an average of 4 different systems (Coveo). 42% of institutional knowledge is unique to one person. When they leave, it leaves. Inefficient knowledge sharing costs large companies an estimated $47M a year (Panopto/YouGov). The average company now runs 101 apps (Okta), and most enterprise data is "dark" — stored but never used, invisible to people and AI alike (IBM). Generic AI chatbots don't fix this: **49% of employees report hitting AI hallucinations at work** (Coveo). An AI that isn't grounded in your documents confidently makes things up. What Ocho changes You don't start fresh — we take what you already have. Ocho connects to SharePoint, Notion, Google Drive, GitHub, and any API with an OpenAPI spec, and keeps syncing on a schedule. Upload the rest by drag-and-drop. We migrate your knowledge for you; nobody retypes anything into a new tool. The brain builds itself. Agentic ingestion analyzes each collection and proposes the right indexing strategy. A knowledge graph of entities and relationships is extracted automatically and updates itself as documents change. You can even measure retrieval quality with built-in evaluations. One shared AI experience, not a per-person tool. Everyone on the team asks the same brain. Answers cite the exact page they came from, so trust doesn't depend on faith. Roles and permissions govern who can change what. An agent harness out of the box. Beyond chat: agentic search, visual graph exploration, and Flows — multi-step automations that loop over your documents, call the AI and external APIs, pause for human approval, and run on schedules or document events. The ROI logic Every answer found in minutes instead of hours claws back part of that lost day a week. Every cited answer is one less hallucination shipped to a customer. And every document ingested is knowledge that stays when its author doesn't. The information already exists — Ocho is how you stop losing it. --- ## Sign in Category: Getting started Description: Sign in with a password or your company's single sign-on. Every Ocho workspace has its own sign-in page at your company's workspace address. The page greets you with "Sign in to" followed by your company's name, and shows your company's logo if one is configured. Sign in 1. Open your workspace address in a browser. 2. Enter your work email in the Email address field and click Continue. Ocho looks up how your organization signs in. 3. Depending on your organization's setup you'll see one of: Password — a Password field appears. Enter your password and click Sign in. Single sign-on — the page says "Your organization (your company) uses single sign-on." and shows a button such as Continue with Google (or Continue with SSO). Click it and sign in with your organization's identity provider; you're returned to Ocho once you're signed in. On either step your email is shown with a Change link that takes you back to the email step. The Email address field starts empty again, so retype the address you want to use. First sign-in with a temporary password If an administrator created your account, the card switches to Set New Password on first sign-in and you must choose a new password before you can continue. The form shows your organization's password requirements as a live checklist ("Password must include") that ticks off each rule as you type, plus a strength meter; it also warns if Caps Lock is on. If you can't get in **"Access to (your organization) is by invitation. Contact your administrator to be added."** — your organization is invite-only. Ask an administrator to add you. "Your account is disabled. Contact your administrator." — an administrator has deactivated your account. **"Your email domain isn't permitted for (your organization). Contact your administrator."** — self-service sign-up is limited to approved email domains. "This account belongs to a different organization." — you signed in with an identity that's linked to another workspace. **"No account found for this email domain. Please contact your administrator."** — the email's domain isn't registered with any workspace. Check the address or contact your administrator. Most of these appear after a single sign-on attempt is rejected. The last one appears at the email step, but only on the general Ocho sign-in page — at your workspace address an unrecognized email simply continues to the password step. Good to know There is no self-service "forgot password" link. If you're locked out, an administrator can reset your password; the new password is emailed to you, and your administrator can also share it with you directly. You can change your own password any time in Settings → My account using the Change password form (requires your current password). When changing your password you can either type your own or use the Generate for me option to create a strong one automatically. For your records, Ocho also emails the new password to you. --- ## Manage users and groups Category: Administration Description: Reset passwords, control table visibility, and share tables with groups. Reset a user's password Who can do this: user administrators (the user_admin role). The Users list only appears if user management is provisioned in your workspace — if there's no Users item in your sidebar, ask your Ocho contact to set it up. 1. Open the user's record from the Users list. 2. Click Reset password. A dialog titled "Reset password for {email}" opens: "Generate a password or type one. It's set immediately and emailed to the user — share it with them directly too." 3. The dialog opens in Generate for me mode with a policy-compliant password already generated — use Copy, Regenerate, or Hide as needed. Switch to Create my own to type a password yourself; in that mode a "Password must include" checklist and a strength meter guide you. 4. Copy the password before you submit — after you click Reset password the dialog closes and the password is not shown again. 5. Submit. The password takes effect immediately and is emailed to the user — you'll see "Password reset — emailed to {email}." If the email can't be sent, the reset still applies and the message tells you to share the password directly. Share a table with groups Who can do this: group administrators (the group_admin role), or the table's owner. Sharing is currently available on Data Manager custom tables. On the table list, a Share button appears when row-level security is enabled for your workspace, or when the table's visibility is set to Restricted. 1. Click Share. The dialog lists the table's current grants (or "Not shared with any group yet."). 2. Pick a group from Select a group…, choose Can view or Can edit, then click the + (Add group) button. 3. Remove a grant any time with its trash button — removal is immediate, with no confirmation. 4. Click Done to close the dialog. Members of a granted group get the chosen access to that table; everyone else's access is unchanged. Control who can see a table Custom tables also have a visibility toggle (the table's owner or a table administrator — the table_admin role — only) with two settings: Company — everyone in your company can see the table. This is the default. Restricted — only you, table admins, and shared groups can see the table. Restricting a table is what makes the Share button appear in workspaces without row-level security. Good to know Groups already granted access are filtered out of the group picker; when every group has a grant you'll see "No more groups to add." In workspaces without row-level security, custom tables default to Company visibility and are visible to the whole organization — no Share button needed. A Restricted table is the exception: it's hidden from the organization and shared only via group grants. Documents, datasets, and chats don't have a Share button today — sharing applies to Data Manager custom tables only. Role assignment itself (giving someone table_admin, for example) is done on the Roles tab of the user's record, and requires the role_admin role — user_admin lets you open user records and reset passwords, but not assign roles. See Roles and permissions. --- ## Block reference Category: Ocho Flows Description: Every workflow block and the variable syntax that connects them. The Blocks sidebar is a single flat list — drag any block onto the canvas and click it to open its inspector. Required config fields are marked with * in the inspector. The blocks below are listed in palette order. Blocks Block · What it does LLM Call — Sends a prompt to an LLM and returns its response as plain text (content). Configure the model, max tokens, and temperature (0–1, lower = more deterministic), plus a system prompt that sets the LLM behavior and persona and a fallback user prompt used when no upstream input is wired. Super RAG Search — Hybrid search over a single dataset; returns the matching chunks (each with a relevance score) and a count. top_k is configurable — valid values are 1–50 (default 10), enforced at run time. Document Full Text — Fetches a document's full extracted text as one blob. Optional max_chars truncates if exceeded; outputs include a truncated flag. Dataset Documents — Lists the documents in a dataset (optional state filter and limit); outputs a documents array and a count. The usual first step of a mass-document-analysis flow, wired into a Loop. Document Chunks — Enumerates a document as an ordered list of addressable chunks in one call — drains all by default, with an optional limit. Each chunk carries heading, page, element type, and reading-order metadata. Chat Skill — Invokes one of your chat skills by its Payload ID (pasted into the skill_id field) and returns its answer (content) with sources. The chat skill brings its own system prompt, RAG configuration, and tool set. Human Gate — Pauses the run (status awaiting_approval) until someone clicks Approve in the run panel; passes its input through unchanged. Anyone who can see the run can approve — there is no Reject button. Response / Passthrough — Passes all inputs through as outputs without transformation. It has no configuration; use it as the terminal node of a flow. Integration Call — Calls an operation on the company's active integration for a source key you type (for example "sharepoint" or "notion"), with parameters entered as raw JSON. If no integration is connected for that source, it returns ok: false with reason not_connected. Scratchpad — Sets run-scoped variables (operations: set, append, extend, increment, append_string) and, via the add_section operation, appends a markdown section (with an optional heading) to the run report. Aggregate — Reduces an array (typically a loop's results) — modes are count, join, concat_text, and collect. Optional separator (default newline) and field to pluck a property from each element first. Loop (for each) — Repeats the blocks inside it for each item of a collection (forEach) or a fixed count (for). continueOnError (default true) keeps going when an item fails; set it to false to stop at the first failure. Variable syntax Pattern · Meaning {{ node_key.field }} — Output of an earlier block. The key is the block's canvas key — auto-generated kebab-case from its display name (dotted sub-paths supported). {{ loop.currentItem }} — The current item inside a loop (dotted sub-paths work, e.g. {{ loop.currentItem.id }}) {{ loop.index }} — The current iteration number (from 0) {{ loop.items }} — The whole collection, inside the loop {{ my_loop.results }} — All iteration results, after the loop finishes. Failed iterations appear as null entries. {{ vars.NAME }} — A variable set by a Scratchpad block Good to know Every functional block on the canvas also shows a red Error output handle, but it is cosmetic today: the runner ignores which handle an edge leaves from, and a block whose upstream block failed is skipped. A branch wired from the Error handle never runs. There is no branching block (condition/router), no custom-code block, and no generic HTTP request block — Integration Call is the only external-call block. LLM Call returns plain text only; it cannot call tools or return structured JSON. Note annotations sometimes appear on saved canvases, but they never execute and there is currently no way to place a new one. The block catalog is fixed platform-wide — admins cannot add, remove, or rename blocks, so the set only changes with an Ocho product release. --- ## Manage records Category: Data Manager Description: Browse a table's records, and create, edit, and delete them in the record view. Browse and search records Opening a table shows its record list: a heading with the table's plural label, the first columns of each record (at most 6, plus an Updated column), and a Search… box. Search… narrows the list as you type. It matches each record's display value only, and only within the rows already loaded (the first 50). Column headers are labels only — there is no click-to-sort and no filter bar. If the table has no records yet, you'll see an empty state such as "No crewss yet." Create a record 1. Click the New button — it's labeled with the table's singular name, for example New Crews. 2. A sheet opens on the right with one field per column. Required fields are marked with a red asterisk (). 3. Fill in the fields and click Save, or Cancel* to discard. Each field uses a control that matches its type: text or number inputs, date and date-time pickers, a switch for yes/no, a dropdown for choice fields, and a record picker for reference fields. Edit a record 1. Click anywhere on a row — a single click opens the full record view. 2. Change any fields in the form. 3. Click Save to apply your changes. Nothing is saved until you do. Use the ← Back to button at the top of the record view (labeled with the table's plural name, for example ← Back to Crewss) to return to the list. Delete a record Open the record and click Delete. Your browser asks "Delete this record?" before it's removed. Records are deleted one at a time — there is no row selection or bulk delete in the list. Activity The record view includes an Activity panel listing one entry per event — the event type (such as created or updated), when it happened, and who acted: USER, AI, AUTOMATION, SYSTEM, or IMPORT. A record with no history shows "No activity yet." Who can do this On company-visible tables, creating and editing records requires one of the roles in the table's write roles list. If that list is empty, nobody can edit the table's records. Deleting records is gated separately by the table's delete roles list — you can have edit rights without delete rights, and the other way around. Bulk operations through Data chat follow the same rules: batch creates and updates need your write role (plus the table's AI-write setting), and bulk deletes need your delete role. Only archiving or restoring a table is limited to the table's admins. Good to know The required-field asterisk is a visual marker only — the form doesn't block submission, but the server rejects records missing mandatory values. Multiple-choice, JSON, and attachment fields are edited as raw JSON text (for example ["a","b"]) in a plain text box. Bulk writes are rate-limited (batches of up to 200 rows, 60 writes per minute), so large batch operations through Data chat take time. --- ## Create and build datasets Category: Documents & datasets Description: Turn a collection of documents into a searchable, chat-ready dataset. Who can do this: Library admins (role key super_rag_dataset_admin). Everyone can view datasets. The Create dataset button is visible to everyone, but creating, attaching documents, and building fail for users without the role. Create a dataset 1. On the Datasets page, click Create dataset. In the dialog, enter a Name (up to 80 characters) and click Create. You land on the new dataset's page. 2. A dataset page has three tabs: Overview, Documents, and QA. The Overview shows a journey stepper: Add documents → Build → Review → Use in chat. 3. Click the stepper's Add documents button — it opens the Add documents from library picker directly, where you select documents. The picker only shows documents not already attached. On the page you land on right after creating, the Documents tab shows "Coming soon" — the full Documents tab, with its Add from library button (visible to Library admins only), is available when you open the dataset from the Datasets list instead. Tell Ocho about the content (optional but recommended) Before building, you can answer a short questionnaire about the dataset — what the documents are, who will query them, what matters. Open it with Tell us about this dataset (on a draft dataset — on the page you land on after creating, the button appears as soon as at least one document is attached; when you open the dataset from the Datasets list, it appears once a document reaches Indexed) or Answer questions (on a live dataset, where it becomes a rebuild path). Ocho uses your answers to tune the build strategy (single-choice answers advance automatically, and a few adaptive follow-ups may appear). At any point from the first question onward, a Skip — just build it for me link at the top of the questionnaire closes it without saving answers. Skipping does not start a build — start it yourself (Build dataset on a draft dataset, Rebuild automatically on a live one). On the summary card at the end you can: Click Save & build with this context to start the build with your answers, or Click Save without building to keep your answers for later. Build Start the build from Build dataset on the Overview (or Save & build with this context in the questionnaire). On the page you land on after creating, the button appears as soon as at least one document is attached — but starting a build before your documents finish indexing fails with an error telling you the documents aren't ready. When you open the dataset from the Datasets list, the button waits until at least one attached document has finished indexing — until then the Overview shows only Add documents. The dataset moves through scanning and building — the stepper shows Analyzing documents… and then Building search index… while the state badge updates. Building takes minutes, depending on volume. Review the strategy The Search setup section on the Overview is always present. Before a build it explains that the strategy is generated when the dataset is built; after a build it shows what Ocho decided: The agent's reasoning, in plain language, under Why these choices. Capabilities cards — depending on the strategy these can include Visual retrieval, Knowledge graph, Scanned-doc OCR, and Context enrichment — each marked On, Recommended, Optional, or Off, with pros and cons. A Technical details disclosure listing the strategy's settings. Refine strategy with AI — a chat where you ask for adjustments ("Disable graph extraction", "Use a larger chunk size"). The AI proposes a Suggested change you Apply or Dismiss. Refinement appears only while a strategy is in the proposed state, and only for Library admins. Use it When the dataset reaches Live, chat and search can query it. Chat works as soon as the build finishes; regular search works too, even without a knowledge graph. Graph features (like the entity graph) are available only after a knowledge graph is built — the graph builds automatically when the dataset goes live. Library admins can opt a dataset out with the Automatic graph updates toggle on the Knowledge Graph card, and can also build or rebuild the graph manually from the same card. On a live dataset the stepper offers Try a new strategy with two options: Rebuild automatically and Answer questions. Rebuilding takes the dataset offline for a few minutes — a confirmation dialog warns you first. Good to know A build needs at least one attached document that has finished indexing — wait for documents to reach Indexed first. If some attached documents aren't indexed yet, the build fails with a clear error telling you to wait or remove them. The document picker shows up to 200 unattached documents. Its search box filters only those loaded documents, so with a very large library some documents may not be reachable from the picker. Answers to the questionnaire are saved (use Save without building if you're not ready to build) — you can revisit and change them before a rebuild via Answer questions. --- ## Explore the knowledge graph Category: Search & knowledge graph Description: Navigate entities, relationships, and their source passages visually. A dataset's knowledge graph is a canvas of nodes — entities Ocho extracted from your documents — connected by relationship edges. You explore it on the Search page. In the Library, a dataset's Overview tab shows a Knowledge Graph summary card (entity, relation, and community counts) with a Search this graph link that opens the Search page's dataset map with that dataset selected — click its cell there to drill into the graph. Not every dataset drills in from the map: the Search page's dataset picker tags some datasets as no graph, and their map cells won't open. The tag governs drilling from the map, not whether a graph exists — a search can still auto-drill into a tagged dataset's graph if one is ready. The dataset map (multiple datasets) Whenever two or more datasets are available and you aren't drilled into one, the Search page shows a dataset map — one region per dataset, sized by document count; datasets outside your current selection appear dimmed. Hover a region for the dataset's name, document count, and last update (connector- backed datasets also show the connector and last sync). On a graph-backed region the tooltip ends with "Double-click to explore this dataset" — click it to drill into that dataset's graph (a single click works too), then use Back to map to return. Regions tagged no graph show "No knowledge graph yet" instead and don't respond to clicks. The dataset picker is hidden while you are drilled into a dataset. Navigate the graph Pan by dragging empty space; zoom with the mouse wheel. Drag a node to move it while you hold it; once you release, the layout takes over again and repositions it. Hover a node to highlight it and its direct connections. Click a node to open the source detail view; click empty space to clear the focused node. To find an entity by name, type it into the main search box — after two characters the suggestions list shows an Entities section. Selecting an entity opens the source detail view on that entity's top passage. Search results on the graph After a search, results overlay document nodes onto the graph — one per source file, labeled with the filename and connected to the entities its hits mention. A fresh search automatically drills into the top hit's dataset and focuses its document node. Click a document node to open that document's best-matching passage for your current query. The source detail view Clicking a node opens a full-screen source detail view: A mini-graph of the surrounding neighborhood, plus a one-line gist of the passage when the passage has a summary (PDF passages often don't — the header then goes straight to the document toolbar). Graph neighbors and Search results tabs — Graph neighbors lists related passages from connected parts of the graph. Chat about this to continue in a chat, and Open in full source doc to read the passage in the original document. Back to graph closes the viewer, but it also exits the drilled dataset — you land back on the dataset map, not the graph canvas. The Previous/Next arrows in the header retrace your hops between sources. Good to know Node click targets are small and the layout keeps moving — a near-miss counts as a background click and clears your focused node instead. Zoom in first if a node is hard to hit. Zooming with a notched scroll wheel moves in small steps; trackpad scrolling zooms more smoothly. While a graph is building or updating, the dataset's Overview tab shows a "Building knowledge graph" or "Updating knowledge graph" indicator on the Knowledge Graph card, with the current stage. An amber Stale badge on the Knowledge Graph card means documents have changed since the last graph build. If automatic graph updates are enabled for the dataset, it updates on its own; otherwise a dataset administrator can use Update graph or Rebuild from scratch — see Knowledge graphs. --- ## Choose data sources Category: Chat Description: Control which datasets the assistant searches when answering. Pick which datasets to search 1. In the composer toolbar, click Sources (the button shows how many datasets are active — for example Sources · 5, or Sources · All). 2. In the dataset map that opens, click a dataset's cell to toggle it on or off. Selected cells show a green checkmark, and the readout below the map tracks your selection (for example "4 of 5 selected"). Use the Zoom in, Zoom out, and Reset view buttons to navigate a crowded map. 3. Send your message — the assistant only searches the datasets you selected. You can't sit in a zero-selected state: deselecting the last dataset resets the panel to All sources (the button reads Sources · All), which searches every dataset. Looking for direct search? Chat always answers through the assistant. To run your query as a direct document search instead — with ranked result cards rather than an AI answer — use the separate Search page (sidebar OCHO → Search). Each result card there shows a relevance score, the document name, page number, and a snippet, with Open source and Chat about this buttons. Good to know When you first load the app, all available datasets are selected. Starting a new chat doesn't reset the selection — it carries over until you change it or reload the page. Hover a dataset's cell in the map to see its name and document count. Connector-backed datasets (like Notion) also show the source connector and a last-synced time; upload-only datasets show an updated date instead — handy when deciding what to include. Datasets that aren't fully built yet don't appear in the Sources picker at all — only live datasets show up. See Create and build datasets. Your selection survives closing the popover and applies to every message you send in any conversation — existing or new — until you change it or reload the page. It's not a per-conversation setting. The Sources popover doubles as the response-settings panel: below the dataset map you'll find the Temperature, Response length, Model, and Mode controls (on mobile the drawer is titled "Sources & Settings"). --- ## What is a knowledge graph? Category: Market comparison Description: How extracted entities and relationships turn documents into a map you can navigate. A knowledge graph is a map of things and how they relate: the people, products, materials, projects, and concepts in your documents (the nodes), and the relationships between them (the edges). Instead of a pile of files, you get a structure you can navigate — "show me everything connected to this alloy," "which projects touch this regulation?" Hand-drawn vs. extracted Most tools that show you a "graph" are visualizing links you made yourself. Obsidian's famous graph view maps the [[wiki-links]] you typed by hand — if you never linked two notes, no connection exists, no matter what the notes say. That's a map of your linking discipline, not your knowledge. An extracted knowledge graph is built by AI reading the document content itself: it identifies entities, infers relationships, merges duplicates ("AI" and "artificial intelligence" become one node), and scores its confidence. Connections you never noticed show up because they're in the text, not because someone remembered to add a link. Why it matters for retrieval Keyword and semantic search find passages that look like your query. A graph answers a different class of question — the connective ones: everything about an entity across hundreds of documents, the chain between two concepts, the clusters in a corpus you've never read end-to-end. Paired with search, the graph also gives AI answers richer context about how facts relate. The payoff is measured, not hypothetical: on entity-rich corpora, graph-based retrieval produces 3.4× more comprehensive answers than vanilla RAG (Microsoft Research, GraphRAG) — and the connective questions are exactly where plain search underperforms. How Ocho does it Automatic. When a search-enabled dataset finishes building, graph extraction runs on its own — entities, relationships, confidence scores, and community clusters. No setup step. Self-maintaining. When documents change, the graph flags itself stale and updates. Chat stays available the whole time. Explorable. Every dataset gets an interactive graph: pan, zoom, click an entity to see its relationships and the exact source excerpts it came from, and jump straight to the cited page. Wired into everything. Search results light up on the graph, citations show a mini-map of related entities, and saved views share a vantage point with your team. See it in action: Explore the knowledge graph. --- ## Find your way around Category: Getting started Description: The home page, sidebar, settings, and where to get help. Home Home is a card grid of every page your company has enabled — one card per page, with the application name shown as the card's subtitle (so one application can produce several cards). Click a card to jump straight to that page. You land on Home the first time you sign in. After that, Ocho remembers where you were: opening the app takes you back to the last page you visited, even on another device. Click Home in the sidebar (or your company name at the top of it) to get back to the card grid at any time. The sidebar The left sidebar is your main navigation: Home at the top, directly under your company name (which also links to Home). Application groups — one labeled group per application, each containing its pages. What you see here depends on your company's subscription tier and your roles; if a page is missing, you may not have access to it. Occasionally a page appears as a card on Home even though it's hidden from your sidebar — that usually means you don't have the role for it, and opening the card will show Access denied. Get help now — opens a dialog where you describe your issue; it's sent to the Ocho team along with your email and the page you're on. Usage — a small Usage indicator with a percentage appears when your workspace has usage limits configured. Click it to open the Usage page and see your current usage against each limit. Settings (gear icon) and your user menu (avatar with Log out). The sidebar collapses to an icon rail — use the toggle in the header. Hover over icons to see labels while collapsed. Settings The gear icon opens the Settings dialog, split into My Settings (personal) and Company Settings (shared, visible if you have permission). The exact categories vary by workspace. Changes save immediately — there's no Save button. My account — see your email and change your password. For your records, we'll also email the new password to you. Preferences — pick a Theme (Light, Dark, or System to match your system preference) and, where available, turn Remember my chats on to save your chat history as long-term memory (it's off by default). Good to know When a new release ships, a persistent notice appears: "A new version is available — Refresh to get the latest features and fixes." Click Refresh to pick it up. If you see "Setup In Progress" instead of the app, your organization hasn't finished onboarding yet; contact the email address shown on that page (or your administrator) to complete Ocho onboarding. --- ## Integrations Category: Administration Description: Connect external services and keep their content synced into Ocho. Integrations pull content from external services (document stores, wikis, repositories) into your Library on a schedule. Custom API connections are different: they register an external API for tooling such as the Flows Integration Call block — they don't sync content. Who can do this: integration administrators create and manage connections. Everyone can view them — without the integration administrator role the Connections page shows the list and connection details but none of the create or manage buttons. Connect a service 1. In the sidebar, open Integrations → Connections and click New connection. 2. Pick the source you want to connect. Two kinds exist: Content sources (SharePoint, Notion, Google Drive, GitHub) — synced services. On the Configure connection step, enter a Connection name, complete any source-specific settings (the step also shows the sync schedule the connection will run on), and click Create & connect. Finish the sign-in that opens in a popup window. Custom API — register any external REST API: Connection name, Base URL, Authentication (None, Basic (username / password), API key, or Bearer token — API key adds Header name and Value prefix fields), and an OpenAPI spec (upload a .json/.yaml/.yml file or paste it) that tells Ocho which operations exist. 3. The Connections list is a table; its Status column shows the sync status (for example Synced). Open a connection to see its verification badge: Verified (with the last-verified date) or Not verified, along with any verification error message. A second Connections tab also lives under Flows ("Manage integration connections used by your flows."). Its connection cards show Verified, Error, or Unverified badges, and it has its own New connection button — but creating connections still requires the integration administrator role. Keep content in sync Content connections sync automatically; the connection page shows Sync status, Last synced, Next run, and Active tiles. (Custom API connections never sync.) From a connection's page, integration administrators can: Sync now — trigger an immediate sync (one at a time; a second request while one runs is rejected). Test connection or Reconnect — verify credentials, or re-run the sign-in popup if verification fails. Delete — removes the connection and its sync history, but synced documents stay in the Library (they're detached from the connection's dataset, not deleted). The confirm dialog says "Synced documents and history will be removed." — in practice the documents survive. Everyone can browse the Sync History and Synced Documents tabs on the connection page. To see a sync finish, refresh the Sync History list — job rows update as they complete. The See full sync history link opens the Sync Jobs page (also in the sidebar under Integrations), pre-filtered to that connection. Bring your own OAuth app (advanced) When creating an OAuth-based content connection (SharePoint, Notion, Google Drive) from the Connections page, the configure step includes a collapsed Advanced: use your own OAuth app section. (GitHub connections use a GitHub App instead, so this section doesn't appear for them.) Enter your organization's Client ID and Client Secret there to use a dedicated OAuth app instead of the platform's shared app; leave the fields blank to use the default. Set this before authorizing — the section isn't available on an existing connection's page. Good to know Synced documents land in the Library like any other document and can be attached to datasets. If a source isn't configured on the platform side, the connection still saves, but a warning appears: "Connection saved, but it can't be authorized yet -- this integration isn't configured on the server. Ask an admin to provision the provider credentials." Custom API connections don't sync — attempting a sync on one is rejected with "Custom API integrations do not sync". --- ## Run and monitor workflows Category: Ocho Flows Description: Execute workflows, watch live progress, approve gates, and read run reports. Run from the editor Click Run in the editor toolbar. Ocho saves any unsaved changes first, then starts the run immediately — there is no compile or validation step and no warnings prompt. A Run started. toast confirms it, and the editor's right-side panel switches from the Assistant to the Run panel with live progress (use the Assistant/Run toggle to switch back). Blocks that can't run don't warn you up front — the block itself fails during the run, and any blocks downstream of it are marked skipped. If other blocks still succeeded, the run ends as a partial failure; if nothing succeeded, it ends as failed. The run panel Status — queued, running, awaiting approval, succeeded, partial failure, or failed. Updates stream in live, typically within a second or two. Run report — every finished run gets a report, even if it's just "ran N steps in Xs". Summary chips sit above the rendered markdown report: Duration and Steps always show, Items appears when the run had loop iterations (shown as succeeded/total), and Errors and Warnings appear only when non-zero. Scratchpad blocks add extra markdown sections to the report. Event timeline — each block's start, success, failure, or pause, in order. Shows "No events yet." until the first event arrives. Approve a Human Gate When a run hits a Human Gate block, its status becomes awaiting approval and an Approve button appears at the top of the run panel. The run resumes once approved. There is no reject button — an unapproved run simply stays paused. Runs history The Runs tab on the Flows page lists recent executions. Each row shows a status badge, the workflow (identified by the first characters of its ID, or Ad-hoc for runs without one), the trigger kind when one was recorded, and the created-at timestamp; failed runs also show a short error summary. Click a run to open its panel in a dialog — report and event timeline, plus an Open in Editor link to the workflow. Ad-hoc runs have no workflow, so no link. Run a workflow proposed in chat Chat can sometimes propose a workflow as a card in the conversation ("Run this multi-step workflow?"). More often it answers with a plain text blueprint instead — a blueprint is not executable. When the proposal card does appear, it offers: View — a read-only preview of the proposed canvas. Edit on Canvas — turns the proposal into a real workflow and opens the editor. Run — also turns the proposal into a real workflow (it appears in your Workflows list), then executes it right there, with the run panel embedded in the card. Proposals never run on their own. Good to know Runs can't be cancelled once started; a stuck run can sit at awaiting approval indefinitely without harm. The history shows the most recent 50 runs. If you refresh the chat page, a proposal card goes back to its original unrun state — it forgets it was ever run or promoted. The run itself continues and remains visible under the Runs tab. Clicking the card's Run again runs the same workflow a second time — Ocho remembers the workflow it already created, so no duplicate appears in your Workflows list. --- ## The AI data assistant Category: Data Manager Description: Ask questions about your tables — and let the AI create, update, and delete records on its own. The Data chat pane on the right side of Data Manager is an AI assistant. Type into the Ask about your data... box to start. Conversation history is kept per table — open the clock icon (Conversations) in the pane header to revisit past chats for the table you have selected. The assistant is not limited to the table you have open: it can work with any table in your workspace that has AI access enabled (and that you can access), and it also answers from the knowledge datasets listed in the pane's Sources picker. The open table only determines which conversation history you see. Ask about your data For tables created with Let AI read this table turned on, ask natural-language questions — "how many open items are assigned to Dana?", "summarize this quarter's entries" — and the assistant can list, search, filter, count, and summarize the table's records. Let the AI change data The assistant can create and update records when both conditions are met: the table was created with Let AI write this table turned on, and you yourself have write rights on the table. Changes are applied directly — there is no approval step, so a request like "close all overdue items" takes effect as soon as the assistant acts. Review the results afterwards in each record's Activity feed. The assistant can also delete records in any table where you have delete rights — deleting needs only your own delete permission, not the table's AI write setting. Who can do this Anyone with access to a table can chat about it, within that table's AI access settings. Only table admins (table_admin) can set AI access — the AI access checkboxes appear in the New table dialog, and per-column visibility is in the Columns sheet. Good to know Let AI read this table and Let AI write this table are both off by default and are set in the New table dialog when the table is created — there is no in-app way to change them afterwards. If the assistant says it can't see or change a table, that's why. Columns can individually be hidden from the AI: untick Visible to AI on a column in the Columns sheet (it's on by default). The assistant may therefore see fewer fields than you do. Review AI-made changes in the record's Activity feed — entries made by the assistant are marked AI. With no table selected, the pane keeps a separate workspace-level conversation history. --- ## Manage datasets Category: Documents & datasets Description: Rebuild, reorganize, and delete datasets safely. Who can do this: users with the Library admin role. Everyone can view datasets. Some edit controls appear for all users, but changes made without the role are rejected by the server. Edit dataset details On the dataset page, click the dataset name to rename it in place, or click Add a description... to edit the description inline. Metadata tiles show Documents, Source, Last updated, Graph updated, and Graph entities at a glance; the graph's Ready/Stale/Failed status badge lives on the Knowledge Graph card below. A Notes section lets you keep a manual record journal — use Add note to start it. Add or remove documents On the Documents tab, Add from library attaches existing documents (never copies them) via the Add documents from library picker. The tab's table is headed Attached documents. Select rows and Detach to remove them from this dataset — the documents stay in the Library and in any other datasets. Rebuild Rebuild after adding or removing significant content, or to change the strategy. On a live dataset the overview shows Chat ready · Try a new strategy: with two buttons: Rebuild automatically opens a Rebuild with a new strategy? confirmation; click Rebuild dataset to start the rebuild right away — no questionnaire. Answer questions opens the Tell us about this dataset questionnaire; answer the questions, then finish with Save & build with this context to rebuild. The Skip — just build it for me link only closes the questionnaire — despite its label, it does not start a build. A live dataset goes offline for a few minutes while it rebuilds — chat and search skip it until it's live again. If a build fails or is cancelled, the journey stepper on the overview shows a Retry build button. Delete a dataset The Dataset actions section on the overview reads: "Delete this dataset only when you no longer need its search setup or graph." Click Delete dataset, type the dataset's name exactly to confirm, then click Permanently delete dataset. This removes the dataset, its document attachments, chunks, and graph data — but Library documents remain available to attach elsewhere. This action cannot be undone. Good to know Datasets are managed individually — the list has row checkboxes, but there are no bulk actions (no bulk delete). The Knowledge Graph card's Rebuild from scratch and Update graph buttons rebuild the graph only — they are separate from the dataset rebuild above. Deleting a document is different from detaching it: delete removes the file everywhere and permanently (see Upload documents). --- ## Citations and sources Category: Chat Description: Verify answers with inline citations and the source viewer. Every grounded answer carries a sources footer — a book icon plus "N sources" and pills naming the datasets the answer drew from (up to four, then "+N"). Many answers also include inline numbered citation markers — small badges like 1, 2 — but only when the model cites specific passages in its text, so don't be surprised if an answer has a sources footer and no inline badges. Preview a citation Hover over a citation marker for a moment to see a preview card: the cited passage (or its summary) and the document name and page (for example, "100m_offers.pdf · p203"). For tables, figures, and charts the card also shows an image preview. The card has exactly two buttons: Open source and Chat about this. Open the source viewer Click a citation badge or an answer's "N sources" footer to open the source viewer — a large dialog that covers most of the screen. (The footer's tooltip says "Open sources sidebar", but it opens this same viewer.) From the Search page, clicking a result card's body does nothing — click the card's Open source button instead. Inside the viewer: The document renders on the right — PDFs open to the cited page with the exact passage highlighted; markdown, text, and transcript files highlight the cited span in place. When the cited chunk has a generated summary, a one-line summary sits above the document. To move around a PDF, use the previous/next page arrows and the editable page-number box (for example, "81 / 206") in the document toolbar — the "Page N of M" pill in the header is a read-only indicator, and the zoom ("100%") and Fit buttons are disabled placeholders marked "Coming soon". The sidebar on the left is tabbed — one tab shows at a time: Citations from answer (every citation from that answer, numbered #1..#N), Graph neighbors (related passages from the knowledge graph), and Search results (a "Search sources…" box). Which tabs appear depends on where you came from — the Citations tab only shows when you arrived from a chat citation. The search box looks across your sources, not just the cited dataset. The mini-graph at the top of the sidebar shows entities related to the passage as unlabeled dots; use its expand button to open the larger Graph neighborhood view, where nodes are labeled. Click a node to hop to that entity's best source passage — the nodes are small, so aim carefully. Open in full source doc (bottom-right, next to Chat about this) opens the complete document on its own page in a new browser tab. The viewer header has Back to chat (or Back to results) and an X to close. Once you've navigated to a second source, Previous source / Next source arrows also appear in the header to step through the sources you've viewed — don't look for them when you first open the viewer; they aren't there yet. Your browser's Back button does the same: it steps backward through viewed sources, then closes the viewer — it won't dump you out of the chat. Chat about a passage Click Chat about this on a citation's hover preview card to pre-fill the composer with the excerpt quoted (this also opens the source viewer — close it to get back to the composer). Nothing sends until you press Enter, so you can add your question first. The Chat about this button inside the source viewer currently closes the viewer without pre-filling the composer — use the hover card's button if you want the excerpt quoted. Good to know Web-result citations (when web search is in play) don't open the source viewer. The sources footer is a single button that always opens the answer's highest-relevance source — you can't pick a specific source from it — so a web source opens in a new tab only when it happens to be the top-ranked one. An inline citation badge for a web result may do nothing when clicked. If a markdown or text document changed after it was cited, the highlight is replaced with a note: "Citation highlight unavailable — the document changed since this citation was indexed. It will resolve once re-indexing completes." Transcripts show the same note with "the transcript changed" instead. The note appears only when you're not searching within the viewer, and PDFs don't show an equivalent note. Ocho drops citations the model invents that don't match a real retrieved source, so numbering always refers to verifiable passages. --- ## What is agentic ingestion? Category: Market comparison Description: Why an AI agent should decide how your documents get indexed — not a fixed pipeline. How documents get into an AI system decides how good every answer out of it will be. Most tools use one fixed pipeline: split every file into chunks of the same size, embed them, done. That works until it doesn't — scanned PDFs come out empty, tables lose their structure, and a legal contract gets chopped by the same rules as a Slack export. Agentic ingestion puts an AI agent in charge of that decision. The agent inspects your actual documents and designs an indexing strategy for them: whether they need OCR, whether tables and figures need visual retrieval, how to chunk, what to enrich, whether a knowledge graph will pay off. How it works in Ocho 1. Tell it what you have (optional). A short questionnaire asks what the documents are, who'll query them, and what matters. Skip it entirely with "just build it for me" if you'd rather not. 2. The agent scans and proposes. After analyzing the content, it presents a strategy in plain language: its reasoning, plus capability cards — visual retrieval, scanned-document OCR, context enrichment, knowledge graph — each marked recommended or not, with trade-offs and cost. 3. Refine it in chat. Disagree? Ask the strategy agent to adjust — "enable visual retrieval," "why is OCR off?" — and accept or dismiss its proposed changes. 4. Prove it works. Built-in evaluations score retrieval quality against golden queries (recall, precision, ranking). If a rebuild regresses, roll back to the previous strategy. Why this beats the alternatives Manual-capture systems (Obsidian, classic "second brain" methods) make you the ingestion pipeline — every note handwritten, every link hand-made. Fixed-pipeline tools (most AI notebooks) treat a scanned blueprint and a markdown memo identically and silently lose whatever doesn't fit. Agentic ingestion adapts to the corpus you actually have — which is the whole point of building on what you already own instead of starting fresh. The numbers back this up: ingestion-side decisions move answer quality more than model choice. Anthropic measured 49% fewer retrieval failures from contextual enrichment at ingestion time alone — one of several build-time choices the strategy agent makes when your corpus calls for it. For the full argument, read the deep dive: What is agentic ingestion? Try it: Create and build datasets. --- ## Usage and billing Category: Administration Description: Understand usage limits, monitor consumption, and manage the subscription. Usage limits Usage budgets run over three window types — Rolling 5h, Weekly, and Monthly — and a workspace can carry any combination of them. A budget is either company-wide or assigned to a specific user. Usage is always shown as a percentage of the limit, never as a currency amount. The sidebar indicator (labeled "Usage") shows your highest current percentage across all windows; it turns amber at 80% and red over the limit. Click it to open the Usage page. The indicator only appears when at least one budget is configured. The Usage page ("Current usage against your workspace limits.") shows one meter per window — labeled Rolling 5h, Weekly, or Monthly — with the percentage and when it resets. When a limit is hit, Chat replies with a Usage limit reached notice that shows the reset time and includes a View usage details link to the Usage page. Workflow runs fail with a usage-limit error message instead. Either a per-user or a company-wide budget can trigger it. Company-wide usage Who can do this: Cost Admins. Cost Admins see a Company Usage section ("Per-user usage across your workspace.") below the personal meters on the Usage page. It's a table with one row per user and one percentage column per budget window. Users are identified by their user ID, not by name or email. If there's no data yet, the section shows "No user usage data available." Billing Who can do this: Payment Admins. Billing runs through a secure external payment partner: During onboarding, the Set Up Billing step ("Complete your subscription setup to activate your account.") is shown to Payment Admins. Clicking Set Up Billing redirects you to the payment partner's checkout; after payment you're returned automatically and the app verifies the subscription before continuing. If you cancel, you'll see "Checkout was cancelled. You can try again when ready." If billing is already active, the step shows Billing Active with a Continue button. Good to know No budgets configured means no meters — the Usage page says "No usage limits are configured for your workspace." and nothing is limited. The sidebar Usage indicator is hidden entirely in this state. Limits reset on their window schedule automatically; there's nothing to clear manually. --- ## Triggers and connections Category: Ocho Flows Description: Start workflows automatically and connect external services. Triggers Triggers are managed from the Triggers button in the editor toolbar. It opens a dialog titled Triggers — "Run this flow automatically when a record changes. Triggers arm on publish." 1. Click Add trigger (the dialog shows "No triggers yet." until you do). 2. Pick a Source: Data Manager record — choose the Table and the On event: Any change, Created, Updated, or Deleted. Library document — choose a Document type: Any, Document, or Transcript. The trigger fires when a matching library document record is created or updated — library document triggers do not fire on delete. 3. Click Save triggers. Saved triggers only start firing once the workflow is published — that is what "Triggers arm on publish" means. The Triggers dialog edits record-event triggers only. Other ways a workflow can start: Manual run — the Run button in the editor toolbar. Schedule, webhook, and chat triggers exist on the platform but have no setup UI yet — they can only be added via the Copilot or the API. For schedules, the cron expression is validated when you publish, and the timezone defaults to UTC unless one is set in the trigger config. Connections The Connections tab manages the external services Integration Call blocks can use. Who can do this: creating and managing connections requires the integration administrator role. Everyone can view them. 1. Click New connection. A Choose a source dialog lists the available sources (Microsoft SharePoint, Notion, Google Drive, GitHub, Custom API — scroll for the full list). 2. Synced sources (SharePoint, Notion, Google Drive, GitHub): the Connect step shows a Connection name field prefilled with the source name — adjust it if you like and click Create. No sign-in happens in this wizard; connect the account afterwards with Reconnect on the connection's detail page (or via the Integrations area). 3. Custom API: the New Custom API connection form asks for a Connection name, the Base URL (the root URL of the API — requests are made relative to it), an Authentication method (None, Basic (username / password), API key, or Bearer token) with the matching credential fields, and an OpenAPI spec you upload or paste (.json, .yaml, .yml) — Ocho reads the available operations from it. Create stays disabled until name, URL, and spec are filled in; after creation the dialog offers a Test connection button to confirm it works. Each connection card shows its verification status — Verified, Error, or Unverified — plus sync info where relevant. Click a card for its detail page: sync status tiles, a See full sync history link, and Sync History / Synced Documents tabs. Integration administrators also see Test connection, Sync now, Reconnect, and Delete buttons there. More detail in Integrations. Good to know Record-event triggers cover Data Manager records and library documents (documents and transcripts) only. Data Manager triggers fire on created, updated, and deleted; library document triggers fire on created and updated only. The Triggers dialog cannot filter on a document state, but a state filter (on_state, e.g. indexed) can be set via the Copilot or API. Triggers arm on publish: editing and saving triggers has no effect until the workflow is published. Scheduled workflows run in UTC unless a timezone was set in the trigger config — check this when results appear at unexpected hours. --- ## Knowledge graphs Category: Documents & datasets Description: How graphs build, stay current, and what rebuilds cost. Each dataset can have a knowledge graph — the entity-and-relationship map that powers the Search page's visual exploration. For exploring a graph, see Explore the knowledge graph; this page covers building and maintaining one. Graph status and controls live in the Knowledge Graph card on the dataset's Overview tab (the dataset page's tabs are Overview / Documents / QA). The card only appears once the dataset has an active strategy — a never-built dataset shows no graph card at all. Who can do this: dataset administrators. Everyone can view and search graphs. Note that the graph buttons and toggle are visible to all users on the dataset Overview — but running them requires the dataset administrator role, and other users' clicks are rejected by the server. Graphs build themselves For search-enabled datasets, the graph builds automatically when the dataset goes live. Chat availability is never blocked by graph work — the dataset stays usable while the graph builds in the background. After that, the graph updates automatically when documents change only if the "Auto-update graph" toggle is on for that dataset. With the toggle off, the card reads "Automatic graph updates disabled — the graph will not update when documents change." and you must run updates manually. Graph states The Knowledge Graph card shows where things stand. A built graph shows a status line like "Graph built 10 days ago from strategy v1." along with Entities, Relations, and Communities counts. State · Meaning No badge — No graph yet — "Knowledge graph builds automatically when the dataset goes live." Building / Updating — Extraction in progress — "chat remains available while this runs" (Updating shows for incremental updates) Ready — Built and searchable Stale — "Documents have changed since the last graph build." Updates automatically only if the auto-update toggle is on Failed — "The automatic graph build did not complete. Use \"Rebuild from scratch\" or \"Update graph\" below to retry." Manual controls Dataset administrators can: Update graph — an incremental update. This button only appears once a graph already exists. The confirmation says "This applies document changes since the last build to the existing graph. The current graph stays live until the update completes." Confirm with Run update. Rebuild from scratch (labelled Build graph if no graph exists yet — in that case the confirmation is titled "Build the knowledge graph?" and confirmed with Yes, build the graph) — wipes and re-extracts everything. The confirmation warns: "Expect roughly $30–150 and minutes-to-hours depending on corpus size." and reassures that "The dataset remains chattable during the rebuild." Confirm with Wipe + rebuild. Use it only when the graph structure itself is wrong. Auto-update graph toggle — the switch at the bottom of the card that turns automatic graph updates on or off for the dataset. There is no way to cancel an extraction once it starts — let it finish. Good to know Graph builds require every attached document to be fully indexed. If some aren't, the confirmation dialog stays open with "Some attached documents aren't indexed yet." and lists the document IDs to wait for before retrying. While extraction runs, the Knowledge Graph card shows the current stage — "Pass 1 — extracting entities", "Pass 2 — extracting relations", "Canonicalizing entities", then "Computing communities" — and refreshes itself every 20 seconds. On the Search page, datasets are picked from a dropdown scope picker (the "All datasets" button with checkbox rows), not pills. The picker's "no graph" tag can lag well behind an actual build — a dataset may still show "no graph" there days after its graph is Ready, even though searching it still surfaces its graph. --- ## Tune responses Category: Chat Description: Adjust creativity, length, and model per message, and switch conversation modes. Per-message settings Click the Sources · N button (the sliders icon) in the composer toolbar. It opens a combined popover: the top half is the dataset source picker, and below it are the response settings. On mobile the same panel opens as a drawer titled Sources & Settings. Temperature (0.0–1.5) — lower is more precise and repeatable, higher is more creative. The readout shows default until you move the slider; once you set a value, a Reset to default link appears below the slider to clear the override. Response length — Default, Short, Medium, Long, or Max. Model — Default, or one of the models your company allows. Choosing Default (or resetting the temperature slider) defers to your workspace's configuration. While any override is active, a small dot appears next to the Sources · N label; it disappears when everything is back to default. Overrides apply to whatever you send next — in any conversation, not just the one where you set them — until you reset them or reload the page. Conversation modes The Mode section is the last section in the same Sources · N popover, beneath Model: Auto — Ocho picks the best mode for each message. Right after a reply, the closed control shows the description of the prompt Ocho picked, with a small auto badge; once you leave the conversation, the control reads plain Auto again. A named mode — each option is shown as a full prompt description with an app suffix (for example "User wants to work with data tables — … · data-manager"). Selecting one pins the conversation to that instruction set; it takes effect from your next message. The switcher is disabled until you've sent the first message, and when your workspace has only one prompt. If the mode is pinned by your chat skill's configuration, the switcher is disabled and shows a lock icon. Good to know Administrators can lock temperature, response length, or model per chat skill; locked values win even if the picker lets you change them. Even unlocked overrides are capped by the server: temperature never exceeds 1.5 (or the skill's own setting, if lower), and response length is limited by the skill's maximum — a picked value may be silently reduced. The model list is controlled by your company — if a model you expect is missing, ask your administrator. --- ## What is agentic RAG search? Category: Market comparison Description: Retrieval that reasons — search as a tool the AI wields, not a single lookup. RAG (retrieval-augmented generation) grounds an AI's answers in your documents: before answering, the system retrieves relevant passages and the model answers from them, citing sources. Grounding is why a RAG answer can be trusted and a bare chatbot answer can't — the model isn't reciting from memory; it's reading your material and showing its work. The stakes are measurable: on Meta's CRAG benchmark (4,409 questions, five domains), frontier models score under 34% accuracy without retrieval, and even industry-best RAG reaches only 63%. Retrieval quality — not model choice — is the gap. Classic RAG has a weakness: it retrieves once. One query, one batch of passages, one answer. If the first search misses — wrong phrasing, question spanning multiple topics, information spread across documents — the model answers anyway, from thin context. Agentic RAG makes retrieval a tool the AI uses deliberately, as many times as the question needs: It rewrites your conversational question into what should actually be searched ("what about the second one?" becomes a real query). It judges its own results — and when they're weak, it broadens the search, reformulates, and retries rather than bluffing. It decomposes multi-part questions into multiple searches and synthesizes across them. It knows when to stop, saying "no confident matches" instead of inventing one. How Ocho does it Ocho's assistant treats document search as a first-class tool. Within a conversation it can search repeatedly, self-correct with an escalation ladder (broaden, then reformulate and retry, then stop), and every retrieved passage carries a citation to the exact page — with the passage highlighted in the source document when you click through. When retrieval genuinely fails, the answer says so instead of hallucinating; hitting an honest "no relevant sources found" is a feature. The same engine backs three surfaces: cited chat, the search page with its knowledge-graph view, and search blocks inside Flows automations — so a workflow can run the same trustworthy retrieval over thousands of documents unattended. Grounding gets you answers you can verify. Agency gets you answers to the questions a single lookup can't reach. You want both. Agentic search is half the story — the other half happens before any question is asked. Read: What is agentic ingestion? --- ## Markdown documents Category: Documents & datasets Description: Write documents directly in Ocho and publish them into search. Besides uploading files, you can author documents inside Ocho with a rich markdown editor, then publish them into the searchable corpus. Who can do this: users with the markdown editor role (document_markdown_editor). Everyone can read published documents. Create a markdown document 1. Open the document library — the Documents card on the Home page (subtitle "Library"); it opens the document library page, heading "Documents", subtitle "Org-wide document library. Upload once, attach to any dataset." Note this is not the sidebar Documents item, which opens the documents table and has no authoring controls. 2. Click New markdown document. In the dialog, enter a Name and click Create draft. (You can also create one from a chat answer with Save to Library.) 3. The draft opens in the editor. New empty documents start in edit mode; existing documents open in read mode with an Edit button. Write The editor supports rich formatting with slash commands — type / for headings, tables, task lists, code blocks, and more (menu items show short aliases, e.g. "Task List · Todo · Check"). Pasting markdown-formatted text from another tool is detected and rendered as rich content automatically. Save, publish, or discard Save draft — keeps your work without touching the published version. Documents with unpublished changes are labeled "Unpublished draft"; never-published drafts also show "Draft (not yet published)". Publish — replaces the searchable version. Clicking it opens a confirmation, "Publish to RAG corpus?" — confirm with Publish. Re-indexing typically takes ~10–30 seconds; the previous version remains searchable until the new one finishes indexing. Publish is disabled while the document is empty ("Add some content before publishing"). Discard draft — appears only while editing and only when you have unpublished changes. It asks "Discard draft?"; confirm with Discard to delete your unpublished changes. The published version, if there is one, is untouched — but a draft that has never been published is deleted entirely. Good to know A draft is not searchable until you publish it. Citations pointing at an older version of the document temporarily lose their exact highlight after you publish; they recover once re-indexing completes. Non-markdown uploads (PDFs and the like) can't be edited in the markdown editor. Uploaded .md files can: their overview shows a "Markdown preview" section — click Edit to write or paste the content, then Save and Publish to make it previewable and refresh the corpus. Re-uploading the original file works too. --- ## Message actions Category: Chat Description: Copy answers, share links to them, save them to the Library, and handle approvals and notices. Hover over an assistant response to reveal its action bar — a row of icon buttons below the answer (underneath its sources row, when the answer cites sources). On touch devices the row is always visible, and a ⋯ (Message actions) button offers the same actions in a menu. Copy and share Copy — puts the full answer on your clipboard and shows a Copied toast. It copies the answer's markdown source (including ** and heading markers), not the rendered text. Copy link — copies a link straight to that message and shows a Link copied toast. Opening the link takes you back to that exact answer, scrolled and highlighted — handy for bookmarking. Conversations are private to you, so the link won't work for anyone else. Save an answer to the Library Who can do this: users with the markdown editor role. 1. Click Save to Library in the action bar. A Save response to Library dialog opens with a suggested title you can edit. The full response is saved as the document body. 2. Click Save. A Saved to Library toast appears with an Open ↗ action — click it to open the draft in a new tab, where you can edit and publish it. See Markdown documents. Tool approvals If the assistant wants to use a tool that needs your sign-off, an approval card appears with Approve and Deny buttons. Approving continues the answer with the tool; denying makes Ocho answer without it. Notices you may see Usage limit reached — your workspace or your own account hit its usage budget (limits can be set at both levels); the notice shows when the limit resets and includes a View usage details link to the Usage page. No relevant sources found in your documents. — document search ran but found nothing confident enough to cite. **Couldn't search your documents right now — answered without them. Try again in a moment.** — search failed, so the answer was written without document context. **Response reached the length limit and was cut off — ask a follow-up to continue.** — the answer hit the length limit; asking a follow-up continues it. stopped — a small badge shown when you stopped the response early. Good to know Copy copies markdown, so pasted text keeps formatting markers like **bold** — paste into a markdown-aware editor for best results. The stopped badge only lasts for your current session. After a reload it disappears, and a response you stopped before any text arrived may show up as a full completed answer. Opening a copied message link scrolls to the message and flashes a highlight on it; the link itself keeps working for the whole conversation. The "no relevant sources" notice doesn't appear on every unanswerable question — sometimes the answer itself just says your documents don't cover the topic. --- ## Ocho vs. Karpathy's second brain Category: Market comparison Description: The viral LLM-wiki pattern proved the idea. Ocho is the product it calls for. In April 2026, Andrej Karpathy posted how he maintains a personal research wiki: raw sources go in a folder, an LLM agent writes and cross-references the wiki pages, an index file keeps it navigable. The gist drew millions of views — deservedly. His core insight is exactly right: "The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping." Let the AI do the bookkeeping, and knowledge compounds instead of evaporating after every chat. Here's the thing: it's a pattern, not a product. Karpathy describes his own tooling as "a hacky collection of scripts" — and adds, "I think there is room here for an incredible new product." What the LLM wiki does well Zero infrastructure: markdown in a git repo, readable anywhere, forever. The LLM maintains cross-references and consistency across pages. Answers get filed back as pages, so knowledge compounds. Where the pattern stops You are the ingestion pipeline. Every source is hand-captured — clipped, pasted, or dropped in a folder. No connectors to where company knowledge actually lives. Nothing syncs; if a source changes, the wiki doesn't know. It's single-player by design. Markdown folders have no permissions, no concurrent-writer safety, no audit trail. A team can't share one. It hits a ceiling fast. The whole approach relies on fitting the index in the model's context window — practitioners report strain around 100–200 articles (one write-up). A company's corpus is thousands of documents on day one. No verifiable citations, no governance, no support. Fine for personal research; disqualifying for anything auditable. What Ocho does differently Ocho is the productized version of the same conviction. The bookkeeping — ingestion strategy, indexing, entity extraction, cross-referencing, freshness — is done by agents, at corpus scale, continuously: Start from what exists. Connect SharePoint, Notion, Google Drive, GitHub, or any API; Ocho ingests and keeps syncing. No clipping. The wiki writes itself — as a graph. Instead of LLM-authored pages, Ocho extracts an actual knowledge graph that updates when documents change. Built for teams. Roles, permissions, usage budgets, and one shared brain the whole company queries — with page-level citations on every answer. No scale ceiling. [Agentic retrieval](/docs/market-comparison/what-is-agentic-rag-search) searches the corpus; nothing needs to fit in a context window. Which should you use? If you're one technical person organizing personal research, run the gist — it's elegant, free, and yours. If you're a team that wants the same compound returns on knowledge you already have, without becoming your own infrastructure, that's the product Karpathy said someone should build. We built it. --- ## Evaluate search quality Category: Documents & datasets Description: Score a dataset's retrieval quality with golden queries and evaluation runs. Each dataset has a QA tab (alongside Overview and Documents) with golden queries and search-quality scores — useful after building, rebuilding, or changing strategy. The QA tab is currently unavailable: it does not load in the app right now and the feature is under review. The sections below describe how it works when it is enabled. Who can do this: running evaluations and managing golden queries requires the Library admin role. Everyone else sees the page read-only. Try queries live The Live search preview at the top of the page lets you probe retrieval against the dataset without writing a golden query. Type a question in the What would a user ask? box and click Search to see the top ranked passages that come back — a quick gut-check without setting anything up. Golden queries Golden queries are saved test questions with expected results. Add the questions your users actually ask: click Add query (Library admins only — the button is hidden for everyone else) to open the Add golden query modal, where you set the query text, a query type (the dropdown lists the raw values data-lookup, comparison, provenance, qa, and procedural), expected keywords, phrases, and pages, ideal document IDs, and an optional expected answer. Each query in the list shows its type as a friendly badge (Data lookup, Comparison, Provenance, Q&A, or Procedural) and a Not yet evaluated badge — per-query pass/fail scores are not yet reported by the eval runner, so this badge currently stays in place even after a run completes. Each query expands to show what retrieval currently returns for it. Admins can also Edit, Approve (for queries awaiting approval), or Delete a golden query from its row. Run an evaluation Evaluations run automatically after a strategy is built. To score the dataset on demand, click Run evals (shown but disabled without the Library admin role); the button is unavailable while a run is already in progress. The results card shows a Pass/Fail verdict badge, the overall score, the number of queries evaluated, and bars for Mean MRR, Page recall, Page precision, Phrase hit rate, and Keyword hit rate. A collapsible Run history panel lists prior runs for the strategy version — select one to view its results. If any queries errored during a run, the page shows how many — per-query error details are not yet exposed by the eval runner, so you can see the count but not which queries failed or why. Roll back a strategy If quality regressed after a rebuild, Roll back (Library admins, available from strategy version 2 onward) switches the dataset back to the prior indexing strategy after a confirmation. Use the View previous strategy versions link to browse eval results from earlier strategy versions for comparison. Good to know Evaluations measure retrieval (did the right passages come back), not the chat model's writing. When the latest run failed or errored, a banner explains what happened with suggested next steps, plus an inline Re-run evals action — and, from strategy version 2 onward, Roll back — for admins. --- ## Ocho vs. Obsidian Category: Market comparison Description: A beloved personal vault vs. a shared brain built from documents you already have. Obsidian is one of the best personal knowledge tools ever made: local markdown files you own outright, linked notes, a famous graph view, and thousands of community plugins. If the app vanished tomorrow, your notes would still open in any text editor. For an individual thinker building a vault of their own writing, it's excellent — and its core app is free. The catch is in that sentence: a vault of their own writing. What Obsidian does well True data ownership. Plain markdown on your disk; no lock-in, works offline, private by default. Flexibility. Thousands of plugins and themes; power users build exactly the workflow they want. Longevity. Your notes will be readable in 20 years regardless of what happens to any vendor. Where it stops You write everything. Obsidian has no ingestion: no connectors for SharePoint, Drive, Notion, or GitHub, no pipeline that reads your existing PDFs and decks. The vault contains what you typed into it. The graph maps your links, not your knowledge. The graph view visualizes [[links]] you created by hand — no entity extraction, no inferred relationships. Past a couple hundred notes it's widely described as beautiful but unnavigable. Single-player. No real-time collaboration, no roles or permissions, no admin controls. Teams bolt on third-party sync tools; it stays a personal tool shared awkwardly. AI is a plugin, per person. Community AI plugins (Copilot, Smart Connections) chat over one personal vault, with limited source attribution and cloud-privacy trade-offs. There's no shared, cited AI over an organization's corpus. What Ocho does differently Ingestion instead of authorship. Connect the tools your company already uses — or drag in files — and Ocho parses, indexes, and keeps syncing. Nobody retypes the knowledge that already exists. An extracted graph. Ocho's [knowledge graph](/docs/market-comparison/what-is-a-knowledge-graph) is built by AI from document content — entities, relationships, confidence scores — and it updates itself when documents change. One brain, whole team. Shared datasets, role-based permissions, and AI chat where every answer cites the exact page — the same experience for everyone, not a per-person vault. A harness, not a plugin. Cited chat, graph-connected search, retrieval evaluations, and Flows automations ship as one product. Which should you use? Both, honestly. Obsidian is superb for your personal thinking — Karpathy's own LLM-wiki setup uses it as a front-end. But the notes you write are a sliver of what your organization knows. For the thousands of documents nobody will ever hand-link — the contracts, specs, decks, and transcripts you already have — you want a brain that builds itself. That's Ocho. --- ## Ocho vs. NotebookLM Category: Market comparison Description: Great for a research project. A different tool than an org-wide, always-current brain. NotebookLM is Google's source-grounded research assistant, and it's genuinely good: upload sources into a notebook, ask questions, get cited answers — plus the famous podcast-style Audio Overviews. The free tier is generous and the grounding quality is real. For a student or a researcher working through a defined set of sources, it's an easy recommendation. The structural difference is one word: notebook. What NotebookLM does well Source grounding with citations — answers come from your uploads, with clickable inline references. Audio and study outputs — Audio/Video Overviews, mind maps, flashcards. Price — a capable free tier; paid tiers from $7.99/month (as of mid-2026). Where it stops Notebooks are silos. Each notebook only knows its own sources — no cross-notebook search, no org-wide brain. Knowledge doesn't respect project lines, but notebooks do. Hard caps per notebook. 50–600 sources depending on tier, 500K words per source (Google's published limits) — and users report retrieval quality dropping as notebooks fill. A company corpus doesn't fit the container. Sources are snapshots. The May 2026 auto-sync covers only Google Docs, Sheets, and Slides (Google). Every PDF, Word file, or non-Google source is frozen at upload — nothing syncs from SharePoint, Notion, GitHub, or anywhere else. You curate every notebook by hand, and collaboration is thin: no roles beyond viewer/editor, no org structure, no workflow automation, no knowledge graph. Enterprise deployment requires Google Cloud. What Ocho does differently One brain, not N notebooks. Datasets are shared, permissioned, and queryable together — ask across everything your company knows, scope down when you want. Continuous ingestion from where knowledge lives. SharePoint, Notion, Google Drive, GitHub, custom APIs — synced on schedules, not re-uploaded by hand. Agentic ingestion picks the right indexing strategy per collection, and built-in evaluations let you measure retrieval quality instead of guessing. A knowledge graph on top. Entities and relationships extracted automatically, kept fresh, explorable visually — something NotebookLM doesn't attempt. An agent harness. The same cited retrieval powers chat, search, and Flows — automations that process documents on schedules and events, with human approval gates. Which should you use? For a bounded research project — a thesis, a briefing, a book club — NotebookLM is excellent and probably free. For a team whose knowledge is large, live, and scattered across systems, the notebook model is the bottleneck: silos, caps, and snapshots. Ocho is built as the opposite: one continuously-synced, self-organizing brain your whole company shares. --- ## Ocho vs. Zaro AI Category: Market comparison Description: A promising new agent workspace — and why a context layer needs a knowledge engine under it. Zaro (zaro.ai) launched from stealth in June 2026 with a $5.1M pre-seed and a pedigree worth respecting — much of the team built Convergence and then Salesforce's Agentforce. Its pitch: a shared "context layer" your AI agents read from and write to, so intelligence accrues to your company instead of each SaaS vendor. "Context compounds. Models commoditise" (TNW). We agree with the thesis — it's ours too. The difference is what sits under the context layer, and how much of it exists today. What Zaro does well The right diagnosis. Agents fail collectively when each tool hoards its own context; a shared layer your company owns is the fix. One-prompt app building. Describe a dashboard or a morning briefing and the workspace generates it; non-technical users edit conversationally. Pricing without seat friction. Credit-based plans from $19/month, unlimited teammates, model-agnostic routing to keep costs down. Where it stops (today) Weeks old. Launched June 2026, pre-seed, eight people, no production enterprise case studies yet. Promising is not proven. Thin ingestion. Confirmed connectors: Gmail, Slack, Notion. No SharePoint, Drive, or GitHub yet, and no document-processing pipeline for the PDF/Office corpus where most company knowledge actually lives. No visible grounding. Public materials describe "recency-weighted retrieval" but show no citations, no source attribution, no knowledge graph, and no retrieval-quality evaluation. When sources conflict, Zaro flags it — good — but you can't trace an answer to a page. Governance is early. Workspace-level siloing, with SSO and audit logs on an enterprise tier; no documented role-based permissions or data residency story. What Ocho does differently A knowledge engine, not just a context bus. [Agentic ingestion](/docs/market-comparison/what-is-agentic-ingestion) processes real document corpora — OCR, tables, visual retrieval — and an extracted knowledge graph keeps itself current. Verifiable answers. Every response cites the exact source page, with the passage highlighted when you click through. Retrieval quality is measurable with built-in evaluations, not asserted. Working automation with guardrails. Flows run loops over documents, call external APIs, pause for human approval, and trigger on schedules and document events — shipping today, with role-based permissions and usage budgets around it. Which should you use? Watch Zaro — the team is real and the thesis is right. But if the job is turning the documents your company already owns into a trustworthy, shared, self-updating brain — with citations you can audit — that's not on their shipping list yet. It's Ocho's core. --- ## Ocho vs. Notion Category: Market comparison Description: Move your work into one workspace — or keep your tools and get the brain anyway. Notion is the standard-bearer for the all-in-one workspace: docs, wikis, databases, and projects in one polished product, now with AI Q&A and connectors on its Business plan ($20/user/month as of mid-2026). If your team lives in Notion and keeps it tidy, Notion AI is genuinely convenient. That sentence carries two conditions, and they're where the comparison lives. What Notion does well All-in-one adoption. Docs, databases, and projects in one place, with a huge template ecosystem and 100M+ users. AI where you write. Inline drafting, autofill, and workspace Q&A with citations, plus connectors to Slack, Drive, Jira, GitHub, and more. Enterprise options. Model choice (GPT/Claude/Gemini) and zero-data- retention arrangements on the Enterprise plan. Where it stops Value depends on living in Notion. Notion AI "is only as accurate as the docs it reads from" (Slite's assessment) — and as organized as your workspace. Messy workspaces get messy answers; the garbage-in problem is yours to manage. Migration is on you. Getting existing knowledge into Notion is manual and lossy — the Confluence importer breaks links and macros, and there's no pipeline that reads your PDF/Office corpus where it lives. Connectors are shallow bridges. They index roughly one year of history, take hours to reflect new content, and skip whole content types (Notion's own docs). The original and the Notion copy drift, and nothing detects it. No knowledge engine. No entity extraction, no knowledge graph, no retrieval tuning or quality evaluation — AI answer quality is a black box you can't measure. Automation is a trigger engine, not multi-step orchestration. What Ocho does differently Keep your tools; get the brain. Ocho ingests from where knowledge already lives — SharePoint, Google Drive, GitHub, and Notion itself — and keeps syncing. Nobody migrates, nobody maintains a second copy by hand. Organization is automatic. [Agentic ingestion](/docs/market-comparison/what-is-agentic-ingestion) designs the indexing per collection, and the knowledge graph extracts structure from content — your answer quality doesn't depend on workspace hygiene. Measurable trust. Page-level citations with exact-passage highlighting, plus built-in retrieval evaluations with golden queries and rollback — you can prove search quality, not hope for it. Deeper automation. Flows loop over documents, call AI and external APIs, pause for human approval, and run on schedules and document events. Which should you use? Keep Notion — it's a great place to write and organize work. That's exactly the point: Ocho doesn't ask you to leave it. Connect Notion as one source among many, and get a cited, self-organizing, measurable brain over everything your company knows — including all the knowledge that was never going to be moved into anyone's workspace. ---