Knowledge base

A knowledge base is a cloud dataset type for unstructured documents. Upload PDF / Word / PPT / Excel / Markdown / text / HTML files; they are parsed, split into chunks and vectorized automatically, then retrieved by the agent alongside your structured datasets.

Create one

Datasets → Add dataset → Create knowledge base. Pick a name, an optional description (it helps the agent decide when to search this KB), the default chunking strategy and the embedding model.

• Structure-aware chunking (recommended) splits along document headings and paragraphs and carries the section path into retrieval — best for manuals and specs.
• By line suits FAQ / list-style documents; fixed length is a fallback for unstructured text.
• Chunk length 200–2000 characters (default 800) with optional overlap.

Files

Upload multiple files at once; each shows parsing → chunking → embedding progress and a per-file retry on failure. Organize files into folders, download originals, and open any file to inspect its chunks — you can re-chunk a single file with different parameters, or edit / delete individual chunks (edited chunks are preserved until you re-chunk).

Scanned PDFs without a text layer are not supported yet.

Recall test

The Recall test tab runs the same hybrid retrieval the agent uses (semantic vectors + keyword matching): tune Top-K, a similarity threshold, and optionally enable rerank for higher precision at some latency cost.

Using it in chat

Mount the knowledge base to any Space just like other datasets. The agent gets a search_knowledge tool and decides when to search documents versus querying tables — a single question can use both (e.g. statistics from SQL plus the handling procedure from your ops manual). Results carry the source file and page so answers stay traceable.

Switching the embedding model

You can change the embedding model later; existing chunks are re-embedded in the background and retrieval keeps working during the switch.