Knowledge

Open Knowledge from the sidebar. The page always edits the globally-selected agent — the one shown in the agent dropdown at the top-left of the header. Switch agents there and the whole workspace re-keys to that agent. Four tabs split the work by source type:

• DocumentsLong-form files — manuals, handbooks, contracts, product spec sheets. We chunk, embed, and retrieve the matching passages on every request. Best for content you already have as a file.
• Web CrawlA public URL (often your help centre, docs site, or product pages). We crawl it, chunk every page, and retrieve passages just like Documents. Best when the content already lives on the web and you want it to stay in sync.
• SnippetsShort, opinionated facts you want the agent to treat as authoritative — pricing rules, brand voice, business hours. Max 300 characters, stored whole (not chunked). Best for anything you can write in a single sentence.
• Q & AA paired question and answer. Both sides are embedded, so retrieval matches whether the user phrases it like your question or like your answer. Best for support-style “how do I…?” content.

Upload a PDF. We extract pages, embed chunks, and the agent retrieves the top-matching passages alongside the user message on every request — automatic RAG, no setup.

• + Add DocumentOpens a file picker. PDFs only. The file is uploaded, queued for OCR/extraction, then chunked + embedded. Status flips from Processing to Ready once indexing finishes.
• Status pillProcessing N%, Ready, or Error. Errors show the reason on hover — common causes are scanned PDFs with no OCR layer and password-protected files.
• SearchFilters the table by filename. Doesn't search inside document content (the agent does that at request time).
• DeleteRemoves the document and all its embeddings. The next request to the agent no longer has access to it.
• Credit indicatorDocument uploads cost credits proportional to total page count. The chip at the bottom shows your running usage. Low-credit accounts see an inline warning above the table.

Point the crawler at a public URL — typically your help centre, docs site, or product catalogue — and we'll fetch the pages, extract clean text, embed each one, and serve them at retrieval time. Re-crawl on demand to pick up changes.

• + Add Web SourceOpens the form. Domain is required; everything else has sensible defaults.
• DomainThe root the crawler starts at (example.com). Use the bare domain — protocol and www are normalised automatically.
• Include pathsComma-separated path prefixes the crawler is allowed to follow. Leave blank to crawl everything under the domain.
• Exclude pathsComma-separated path prefixes to skip. Useful for keeping /admin, /account, or noisy archive sections out of the index.
• Max depth / Max pages / ConcurrencyBounds the crawl. Defaults are 3 / 100 / 4 — enough for most product help centres without burning credits.
• Page statusPer-page sub-status visible when you expand a source: ready, processing, error, skipped (robots.txt or non-HTML).
• Re-crawlPer-source action. Re-fetches every page, updates embeddings, and replaces stale content.
• DeleteRemoves the source and all its indexed pages. Reversible only by re-adding and re-crawling.

Short facts the agent should treat as authoritative. Unlike Documents, snippets aren't chunked — each one is stored as a complete unit and retrieved when its embedding is most similar to the user's question.

• + Add SnippetOpens the inline composer. Type up to 300 characters and save. Saved snippets start retrieving on the next request.
• Character counterLive counter under the textarea — turns red near the cap and blocks save at 300. Keeping snippets short forces precise wording, which retrieves better.
• SearchFilters the table by snippet text. Doesn't affect retrieval — purely a UI filter.
• DeleteRemoves the snippet. Reversible only by re-typing.

A Q&A entry is a question paired with the answer you want the agent to give. Both the question and the answer are embedded separately, so retrieval fires whether the user phrases their message like your question ("how do I reset my password?") or like your answer ("password reset link"). It's the best fit for support-style content where the same fact gets asked in many different ways.

• + Add Q&AOpens the inline composer with a Question field (max 200 chars) and an Answer field (max 500 chars). Save embeds both halves into the agent's knowledge base.
• Status pillProcessing while the pair is being embedded, then Ready. An Error means embedding failed — delete and re-add.
• SearchFilters the table by question or answer text. Purely a UI filter — retrieval always searches the full set.
• DeleteRemoves the pair and both embeddings. The agent stops retrieving it on the next request.

Tools live in their own MCP sidebar item, separate from Knowledge. Connect external tools and live data via the Model Context Protocol — the agent discovers the available tools automatically and you choose which to expose. See MCP Servers for transports (stdio / SSE / HTTP), auth, and tool selection.

On every API call, the agent embeds the user's message and looks up the top-K most similar chunks across Documents, Web Crawl pages, Snippets, and Q&A pairs combined. Matching passages are injected into the prompt before the LLM runs. You don't configure any of this — it's automatic and tuned per agent.

Open Tracing on any request to see exactly which chunks were retrieved, their similarity scores, and which source they came from. That's the fastest way to debug "why didn't the agent know X?".