Knowledge Base
Understand document processing, chunking, and hybrid vector and full-text retrieval
A Knowledge Base converts documents into chunks that Agents and Workflows can retrieve. The resource belongs to a team; you select an embedding model when creating it, and optionally a rerank model.
Processing pipeline
After upload or URL import, a document starts in Pending. Processing performs text extraction, cleaning, chunking, embedding, and indexing, ending in Completed or Failed. The document-level states are pending, processing, completed, error. Chunking uses deterministic character-level splitting (RecursiveCharacterTextSplitter); semantic/ML chunking is not supported. Vectors are stored in a dedicated Qdrant collection per embedding dimension ({prefix}_{dimension}), determined by the selected embedding model.
Chunks of a processed document can be viewed, edited, added, and deleted. Re-chunking replaces all existing chunks and vectors, and manual edits are lost.
Three retrieval methods
| Method | When to use |
|---|---|
| Vector retrieval | Recalls by semantic similarity |
| Full-text retrieval | Matches exactly by lexical/BM25 |
| Hybrid retrieval (default) | Fuses vector and full-text rankings with weighted RRF |
Hybrid retrieval defaults both the dense weight and the lexical weight to 1.0, with rrf_k at 60. The two weights cannot both be 0. Rerank can re-score candidate chunks after fusion.
Retrieval lab
Hit testing / retrieval lab runs single-configuration tests or A/B comparisons, showing each result's rank and score at the vector, full-text, fusion, and rerank stages, and can apply local presets to production Knowledge Base settings.
Related content: Import documents, Chunking, Knowledge Base settings reference.
How is this guide?