Knowledge Base Guide
Create knowledge bases, import documents, tune chunking, and verify retrieval quality
A knowledge base stores a searchable document collection for a team: raw files or web pages are parsed into text, split into chunks, and converted into vectors by the team's embedding model. Retrieval recalls chunks by similarity and hands them to an Agent or workflow as context. This section follows that order across five steps; for the field-by-field reference see the Knowledge Base settings reference.
Prerequisites
Your team must be authorized for one embedding model before you can create a searchable knowledge base, and for a rerank model before reranking can do anything. Authorization and provider setup are covered in Managing Models; an unauthorized model is rejected at creation or retrieval time.
Create a Knowledge Base
Import Documents
Browsing Documents & Metadata
Chunking and Processing
Retrieval and Quality Tuning
Key defaults
| Item | Default | Notes |
|---|---|---|
Chunk size chunk_size | 1000 characters | Minimum 100; characters, not tokens |
Chunk overlap chunk_overlap | 100 characters | Minimum 0 |
Rerank rerank_enabled | true | Has no effect without a bound rerank model |
Rerank candidate pool rerank_candidate_k | 10 | Range 1-100 |
Search request search_mode | hybrid | The KB setting itself defaults to null |
Search request top_k | 5 | Request range 1-20 |
| Single-file upload limit | 50MB | Site setting kb_document_max_upload_size_mb, range 1-1024 |
Lifecycle of one document
Upload/import → pending → preview chunks (optional) → process → extract text → clean → chunk → embed → lexical index → completedOnly completed documents take part in retrieval. pending only means the file is stored — it is not indexed yet; see Import Documents.
For tuning guidance (chunk granularity, weights, thresholds) see Knowledge Base Optimization, and for the conceptual ingest/retrieval flow see Knowledge Base Concepts.
How is this guide?