Knowledge Base Settings Reference

Look up chunking, retrieval, hybrid search, rerank, and document processing fields

Knowledge Base settings affect both document processing and runtime retrieval. After creation, the embedding model and vector dimensions should stay stable.

Embedding Model & Dimensions

Field	Type	Default	Limit/Description
`embedding_model_id`	UUID/null	`null`	Embedding model bound to the Knowledge Base; immutable after creation
`embedding_dimension`	integer/null	`null`	Detected and locked on first vectorization; later vectors must match exactly

When a Knowledge Base is created, embedding_dimension is empty; the system writes the measured dimension after vectorizing the first batch of document chunks. Afterwards, updating embedding_model_id is rejected, and submitting vectors with a mismatched dimension raises DimensionMismatchError. For provider coverage and Qdrant multi-dimension collection mapping, see Embedding Adapters & Vector Dimensions.

A Knowledge Base's visibility defaults to private (visible only to the creator) and can be set to team (visible to team members); public is retained for compatibility only.

Chunking & Rerank

Field	Type	Default	Limit
`chunk_size`	integer	`1000`	`>=100` characters
`chunk_overlap`	integer	`100`	`>=0` characters
`separator`	string/null	`null`	Custom text separator
`rerank_model_id`	UUID/null	`null`	Team-authorized rerank model ID; used together with `rerank_enabled`
`rerank_enabled`	boolean	`true`	Whether rerank is enabled
`rerank_candidate_k`	integer	`10`	`1-100`
`rerank_score_threshold`	number/null	`null`	`0-1`

Retrieval

Field	Type	Default	Limit/Description
`search_mode`	`vector`, `fulltext` or `hybrid`	`null` (Knowledge Base settings); request default `hybrid`	Hybrid fuses vector and full-text search
`top_k`	integer/null	`null` (settings); request default `5`	Request range `1-20`; settings range `1-100`
`score_threshold`	number/null	`null` (settings); request default `0`	`0-1`
`dense_weight`	number/null	`null` (settings); request default `1.0`	`>=0`
`lexical_weight`	number/null	`null` (settings); request default `1.0`	`>=0`
`rrf_k`	integer/null	`null` (settings); request default `60`	`1-1000`

Hybrid retrieval requires at least one of dense weight or lexical weight to be greater than 0. Query length is 1-1000 characters; batch comparison supports at most 10 configurations and IDs must be unique.

expand_adjacent is an internal option

The retrieval layer also has an expand_adjacent switch (default false): when enabled, it pulls in the chunks adjacent to a matched chunk (index ±1) within the same document to round out the context. It is an internal request parameter of the retrieval service and is not currently exposed via the public search API, so it cannot be configured in Knowledge Base settings or in a search request; the public search endpoint does not accept the field.

Documents & Statuses

Supported formats: PDF, DOC/DOCX, TXT, Markdown, HTML, CSV, XLS/XLSX, PPT/PPTX, JSON, and URL. Document statuses are pending, processing, completed, error. The site-wide default single-file upload limit is 50MB and can be changed within 1-1024MB.

Failure Diagnostics

Retrieval responses may include inactive, missing_embedding_model, timeout, failed diagnostics as well as recall, rerank, context, total stage timings. For production troubleshooting, check in order: document completion status, embedding dimensions, Qdrant, PostgreSQL/pg_search, rerank model credentials, and provider quotas.