Documentation
Knowledge bases
Your documents, prepared for retrieval, scoped to a department, with a sensitive tier.
What a knowledge base is
A knowledge base belongs to a department and holds documents: policies, manuals, circulars, whatever the work needs to consult. Agents search it; they are never given it wholesale.
How a document is prepared
- The document is parsed, including tables and layout, not only its plain text.
- It is split into passages sized for retrieval.
- Each passage is enriched with the context it came from, so a passage still makes sense on its own.
- Passages are embedded and stored with the document they came from.
How retrieval works
A search combines meaning-based and keyword matching, merges the two rankings, and can rerank the result. An agent gets passages and the documents they came from, so an answer can always be traced back to a source.
Who can reach it
- The knowledge base must belong to a department the agent's workflow runs in.
- It must be attached to that agent.
- If it is marked sensitive, publishing a workflow that reaches it requires explicit authorization, and the builder cannot read it while authoring.
Mark a knowledge base sensitive when its contents should never be visible during authoring, only at run time. Client files and personal data are the usual cases.