Working with knowledge bases

A knowledge base is an organised library of topics with a private vector index behind it. The index lives on your disk — everything you save is chunked and embedded locally, and nothing uploads.

One memory per job

Make a KB per client, project or domain rather than one big pile. Scope is the point: a chat attached to a KB answers from that KB alone, so client work never leaks into policy work, and retrieval stays sharp because the index only holds what belongs.

Search by meaning

The index is semantic. Ask about "my daughter's creative interests" and it finds the note that says she likes drawing — related ideas surface even when the wording is completely different. Images are part of the same index: diagrams and figures from imported documents come back through the same search as the text.

Scoped or universal

By default you search the KB you're in. When you can't remember where something lives, search everything — one query sweeps every knowledge base at once and tells you where each hit came from.

Using a KB in chats

Attach a KB to a chat and retrieval pulls relevant passages per question; load content when the model should hold it whole. The two gears are covered in Working with chats.

Organising

KBs are cheap to create and easy to curate: drag topics into them, between them, and into the order that makes sense. A KB grows from a handful of notes into a documentation hub without changing shape.

Exporting a library

A whole KB exports as browsable HTML — every topic as a page, with an index to navigate by. Hand the folder to a colleague, drop it on a shared drive, or archive a finished project whole. Individual topics export as HTML or PDF too; see Working with topics.

Durability

Nothing in a KB expires. The page you captured in March still answers in December, even after the live site changed or disappeared — your copy, with its capture date, is the record.