Knowledge bases

Upload documents; agents answer from them. A knowledge base is the "answer based on our docs" primitive: retrieval-augmented generation without any plumbing.

Uploading

In Settings → data → knowledge bases, create one, then drag in documents: .pdf, .txt, .md, .csv, .json (up to 20 MB each, five at a time). Each document is:

1. read: PDFs are transcribed with AI vision, so scanned documents work;
2. chunked: split into overlapping ~800-token passages along paragraph boundaries;
3. embedded: indexed for semantic search, asynchronously (chunks show a pending dot until indexed).

Chunk count, a retrieval test box, and the newest chunks with their source documents.

Filling a base from a workflow

The save to knowledge step writes a document from inside a run: index a repository on push, keep a docs site searchable, file what an agent learned. The source name is the document's identity (a path, a URL), so re-runs replace instead of duplicating; a delete mode removes one. An agent step can also get a base as a tool and save or delete documents itself. Over the REST API, POST and DELETE on /knowledge-bases/:id/documents do the same from outside.

Retrieval

Bind the base on an agent step. Each run, the agent embeds its prompt, retrieves the most relevant chunks (default 5, set with the Retrieved chunks slider), and answers with them as context. The agent is instructed to say so when the answer isn't in the retrieved material, rather than invent one.

The search box on the base's page runs the same retrieval the agent runs. It's the fastest way to check "would the agent find the refund policy?" before wiring anything.

Gotchas

• Retrieval is top-K, not whole-corpus: the agent sees the K most relevant chunks, not everything. Questions whose answer is smeared across many documents ("summarize all 40 PDFs") aren't a retrieval fit; that's a loop over documents.
• Uploads are per-base. Deleting and re-uploading a revised document is the update path.
• Structured, keyed facts (prices, fees, per-county values) belong in a store, because exact lookups beat semantic search for exact data.