Fine-tune how Opentrace retrieves and processes information from your documents. These settings are configured per-project in the Knowledge Base > Settings tab.
| Setting | Type | Default | Description |
|---|---|---|---|
rag_strategy | Enum | basic | Retrieval strategy: basic, hybrid, multi_query_vector, or multi_query_hybrid. See RAG Strategies. |
agent_type | Enum | simple | Agent type: simple (RAG only) or agentic (supervisor with web search). See Agent Types. |
embedding_model | String | text-embedding-3-large | OpenAI embedding model used for vectorization. Determines the embedding dimensions. |
num_chunks | Integer | 5 | Number of chunks to retrieve per search query. |
final_context_size | Integer | 5 | Number of top chunks to include in the final LLM context after ranking. |
similarity_threshold | Float | 0.3 | Minimum cosine similarity score for a chunk to be considered relevant. Range: 0.0 – 1.0. |
use_reranking | Boolean | false | Whether to rerank retrieved chunks using a cross-encoder model before final selection. |
multi_query_count | Integer | 3 | Number of query variations to generate (only for multi-query strategies). |
vector_weight | Float | 0.7 | Weight of vector similarity in hybrid search RRF fusion. Range: 0.0 – 1.0. |
keyword_weight | Float | 0.3 | Weight of keyword (full-text) search in hybrid search RRF fusion. Range: 0.0 – 1.0. |
Start simple, then optimise: Begin with defaults (Basic strategy, 5 chunks). Only increase complexity if you're not getting the answers you expect. Multi-query hybrid with reranking gives the best results but is the slowest and most expensive.
num_chunks if…similarity_threshold if…