Skip to main content
Knowledge bases hold documents that are chunked and embedded so agents can retrieve grounded context.

​
How it works

  1. Create a knowledge base.
  2. Add documents. Each document is chunked (~1000 chars) and every chunk is embedded.
  3. Retrieval runs a cosine-similarity search over the chunks with pgvector.
Embeddings are produced by a self-hosted embedder (bge-small-en-v1.5, 384-dim), so no external embeddings key is required.

​
Supported formats

Up to 4MB per file. Text-family files are decoded immediately. Everything else is queued and converted by a background worker, so an upload returns as soon as the document is accepted rather than blocking on conversion - a large PDF becomes searchable a little after it is uploaded, not instantly.

​
Documents become Markdown, not flat text

Conversion is done by anydoc, which renders every format through one Markdown serializer. Headings, tables, lists, and speaker notes survive. That matters because a chunk is retrieved on its own, with no surrounding document to explain it. A heading tells the reader what the chunk is about, and a table keeps each figure next to the column that names it - | Q2 | 14 | under a Quarter | Widgets header answers a question that a bare Q2 14 cannot.
A PDF that is a scan or a photo has no text to convert. Those uploads fail with an explanation rather than embedding an empty document; OCR is not currently part of the pipeline.

​
From the CLI

​
From an agent

Add a tool with the built-in key knowledge_search and reference it from an agent node. At run time the agent calls knowledge_search(query), which embeds the query and returns the top matching chunks for the model to answer from.

​
Knowledge graphs

A base can also extract an entity-relationship knowledge graph from its documents, so agents retrieve related facts by traversing the graph, not just by similarity. See Knowledge graphs (GraphRAG).

​
API

  • POST /v1/knowledge - create a base
  • POST /v1/knowledge/{kb}/documents - ingest a document
  • POST /v1/knowledge/search - search chunks
See the API overview.