hybrid.yaml

# This example demonstrates:
# 1. Running multiple retrieval strategies in a single RAG source
# 2. Shared docs across all strategies
# 3. Strategy-specific additional docs
# 4. Per-strategy limits for fusion input
# 5. Final limit after fusion
# 6. Full hybrid retrieval with RRF

agents:
  root:
    model: openai/gpt-5-mini
    description: assistant with hybrid search
    instruction: |
      You are a helpful assistant with access to hybrid retrieval
      combining semantic and keyword search for comprehensive results.
    toolsets:
      - type: rag
        ref: knowledge_base

rag:
  knowledge_base:
    tool: 
      description: to be used to search for information about blorks
    docs:
      - ./blork_field_guide.txt   # Shared across both strategies
    strategies:
      # Chunked embeddings strategy for semantic search
      - type: chunked-embeddings
        embedding_model: openai/text-embedding-3-small
        docs:
          - ./docs                            # Additional docs for chunked-embeddings strategy only
        database: ./chunked_embeddings.db     # chunked-embeddings database to use; path to local sqlite database
        similarity_metric: cosine_similarity  # similarity metric (default: cosine_similarity)
        threshold: 0.5                        # cosine similarity threshold (0-1, default: 0.5)
        limit: 20                             # Chunked embeddings retrieves top 20 candidates
        batch_size: 100                       # chunks per embedding API call (default: 50, max: 2048 for OpenAI)
        max_embedding_concurrency: 5          # concurrent embedding batch requests (default: 5)
        vector_dimensions: 1536               # vector dimensions of the model in use
        chunking:
          size: 1000
          overlap: 100
          respect_word_boundaries: true
      # BM25 strategy for keyword matching
      - type: bm25
        docs:
          - ./docs                            # Additional docs for BM25 strategy only
        database: ./bm25.db # where to store the BM25 database (path to local sqlite database)
        k1: 1.5             # term frequency saturation (default: 1.5)
        b: 0.75             # length normalization (default: 0.75)
        threshold: 0.3      # BM25 score threshold (default: 0.0)
        limit: 15           # Get the top 15 candidates
        chunking:
          size: 1000
          overlap: 100
          respect_word_boundaries: true
    # Results fusion and limiting
    results:
      fusion:                          # Configures how to combine the results from the different strategies
        strategy: rrf                  # Fuses up to 20+15=35 total candidates with RRF (Reciprocal Rank Fusion)
        k: 60                          # RRF smoothing parameter
      deduplicate: true                # remove duplicates from the results
      limit: 5                         # Final: return top 5 after fusion
      # return_full_content: true        # Return full document instead of chunks (default: false)