Knowledge Base
Centralize your team's documents and web content, then let the AI search, retrieve, and cite them automatically. Knowledge lives in folders you organize, share, and attach to agents โ and it's the foundation of every grounded AI answer.
How it works
When you add a document, YOffice runs it through a four-stage pipeline that makes its content searchable by the AI.
Upload a file or link a website. Each source is processed independently.
Text is extracted from the file, then split into sized chunks that respect paragraph and sentence boundaries.
Each chunk is turned into a vector embedding and written to pgvector alongside its text.
When you ask a question, the closest chunks are retrieved and cited in the answer.
Supported file types
YOffice extracts text from a broad range of document formats. Spreadsheets are converted into an LLM-friendly structured form (and made queryable), so the AI can reason over tabular data, not just prose.
The converter reads the text already present in a file, so text-based documents give the most reliable retrieval. For a scanned or photographed page (an image-only PDF), supply a text-based version to make its content fully searchable.
Semantic retrieval with pgvector
Chunks are embedded and stored in pgvector. At query time YOffice searches by vector similarity (cosine) to find the most relevant chunks, falling back to keyword matching if a vector search comes back empty, and the AI cites the results inline. Each document has an AI-context toggle, so you control exactly what the AI may draw on.
| Embedding model | Provider |
|---|---|
| text-embedding-3-small | OpenAI (default) |
| text-embedding-004 | |
| nomic-embed-text | Ollama / Ollama Cloud |
Different embedding models produce vectors of different dimensions, so YOffice doesn't silently fail over between providers โ your organization's content is embedded with one model across every folder. Pick your embedding provider deliberately at the organization level.
Folders, sharing, and limits
Folders are how you organize and share knowledge. Each folder supports:
- Four sharing scopes โ Private, Specific people & Workers, Workspaces, and Organisation โ each with view or edit permission. See Members & Roles.
- A 1,000-document limit per folder to keep retrieval fast and relevant.
- Folder watchers โ point a watcher at a local folder and new files are ingested automatically as they arrive.
- Open Knowledge
Click Knowledge in the navigation rail. The Knowledge tab lists your folders.
- Create a folder
Click New Folder and give it a descriptive name like "Engineering Runbooks".
- Add sources
Upload files or link a website. Each source processes independently โ a large PDF won't block smaller ones.
- Set sharing
Choose a sharing scope and permission. You can change it any time.
Website as a knowledge source
Beyond files, you can link a website as an AI source. YOffice crawls the site within configurable caps, generates an AI summary at ingest, and indexes the content according to the RAG mode you choose.
Always-on AI summary
At ingest, YOffice produces a real AI summary of the site using a map-reduce pass over the page chunks โ it summarises chunks, then summarises the summaries. This is a genuine model-generated overview, not a naive truncation, so even in the lightest mode the AI has a faithful picture of the source.
Three RAG modes
The mode controls what gets indexed for retrieval:
| Mode | What's indexed | Use whenโฆ |
|---|---|---|
| SummaryDefault | The AI summary only | You want a compact, fast overview of the site. |
| Summary + full | The summary and the full page text | You want the overview plus the ability to retrieve specific details. |
| Full | The full crawled text | You need maximum recall over the entire site's content. |
Crawl caps
Crawling is bounded by validated limits โ maximum pages, maximum bytes per page, and crawl depth. Each has a hard positive minimum so a website source can't run away with your budget.
Using knowledge in AI chat
Once a folder is indexed, the AI can draw on it in two ways:
- Scope per chat โ open the Knowledge tool in the chat tool strip and pick folders for that session, or add a folder with #.
- Always-on knowledge โ every folder you can see (owned, shared with you, your workspaces, or organization-wide) is searched automatically across sessions, and each document's own AI-context toggle lets you exclude specific files, ideal for a company wiki or product docs.
Attaching folders to agents
Attach one or more knowledge folders to an agent and that agent retrieves from them automatically when it runs โ a support agent linked to your "Support Docs" folder answers questions without any manual lookup. Folders can likewise be shared with an agent.
Organize by topic or team rather than by date. Use descriptive folder names. Keep content current โ outdated documents mislead the AI. For external docs that change often, link the website source so the content stays fresh, and pick the RAG mode that matches how much detail you actually need to retrieve.