Knowledge bases
Retrieval settings
Passages, relevance threshold, meaning/wording weight and reranker — and which take effect immediately.
The Settings button on a knowledge base opens a two-part dialog: Retrieval and
Chunking. Needs knowledgeBase.update.
Retrieval settings change the very next answer. Chunking settings only reach documents that are re-indexed after the change. Retrieval edits are done when saved; chunking edits need a Re-index of every document you want them applied to.
Retrieval
| Field | Default | Meaning |
|---|---|---|
| Passages (top K) | 6 | How many passages the model receives to write the answer. More = broader coverage, more credits, more noise. |
| Min. relevance | 0.2 | Passages scoring below this are dropped. Raise it when answers keep pulling in unrelated passages. |
| Meaning weight | 0.7 | The balance between meaning (vector) and wording (keyword) search in Hybrid mode. 0.7 means meaning decides and wording breaks ties. Lower it when people search by codes, proper names or exact terms. |
| Reranker | None | A model reads the question together with each candidate passage and reorders them. A second pass that reads the question and each candidate passage together. More precise, costs an extra rerank operation. Only shown when the workspace has a rerank model (Cohere, Voyage, Jina). |
People can override Passages and the search mode per conversation — see Scope and search mode.
Chunking
The same fields as when creating the knowledge base: method, keywords, questions, Summary tree. Changes here become the default for files uploaded later and for files you Re-index.