config.yaml

# Semantic Search Engine Configuration
# =====================================

# Model Configuration
models:
  embedding:
    name: "text-embedding-3-large"
    provider: "openai"

  chat:
    name: "gpt-4o-mini"
    provider: "openai"
    temperature: 0.0

# Document Processing Configuration
document_processing:
  chunk_size: 1000
  chunk_overlap: 200
  add_start_index: true

# Vector Store Configuration
vector_store:
  provider: "chroma"
  collection_name: "semantic_search_docs_streamlit"
  persist_directory: "./chroma/db"

  # ChromaDB connection mode
  # Set use_docker: true to connect to ChromaDB Docker container
  # Run: docker run -d -p 8000:8000 chromadb/chroma
  use_docker: true
  chroma_host: "localhost"
  chroma_port: 8000

  # Retrieval settings
  search_type: "similarity"
  search_k: 3  # Number of chunks to retrieve

# Hybrid Retrieval Configuration
hybrid_retrieval:
  # Enable hybrid search (BM25 + semantic)
  enabled: true

  # Retrieval method: "semantic", "bm25", "hybrid"
  default_method: "hybrid"

  # Weight for semantic search in hybrid mode (0-1)
  # 0 = BM25 only, 1 = semantic only, 0.5 = equal weight
  alpha: 0.5

  # RRF constant for rank fusion (typically 60)
  rrf_k: 60

  # BM25 parameters
  bm25:
    k1: 1.5  # Term frequency saturation
    b: 0.75  # Length normalization

  # Re-ranking configuration
  reranking:
    enabled: true
    # Provider options:
    #   - "auto": tries jina (local) first, then cohere (cloud)
    #   - "jina": force local Jina model (requires sentence-transformers)
    #   - "cohere": force cloud Cohere API (requires COHERE_API_KEY)
    provider: "auto"

    # Cohere model (if using cohere)
    cohere_model: "rerank-english-v3.0"

    # Jina model (if using jina)
    jina_model: "jinaai/jina-reranker-v1-tiny-en"

    # Number of candidates to fetch before reranking
    fetch_k_multiplier: 3

# Retrieval Presets Configuration
# Pre-configured profiles for different use cases
retrieval_presets:
  high_precision:
    display_name: "High Precision"
    description: "Fewer results, higher relevance. Best for specific questions."
    icon: "🎯"
    k: 3
    alpha: 0.7
    rerank: true
    method: "hybrid"

  balanced:
    display_name: "Balanced"
    description: "Good balance of precision and coverage. Recommended default."
    icon: "⚖️"
    k: 5
    alpha: 0.5
    rerank: true
    method: "hybrid"

  high_recall:
    display_name: "High Recall"
    description: "More results, broader coverage. Best for exploration."
    icon: "🔍"
    k: 10
    alpha: 0.3
    rerank: false
    method: "hybrid"

  # Default preset to use
  default_preset: "balanced"

# Conversation History Configuration
conversation:
  # Enable conversation history
  enabled: true

  # Storage directory for conversation JSON files
  storage_dir: "./conversation_history"

  # Maximum queries to retain per session
  max_history: 50

  # Enable follow-up question optimization
  follow_up_optimization: true

  # Number of recent Q&A pairs to use for follow-up context
  context_window: 3

# A/B Testing Configuration
ab_testing:
  # Enable A/B testing
  enabled: true

  # Storage directory for experiment results
  storage_dir: "./ab_testing_results"

  # Default variants to compare
  default_variants:
    - "semantic"
    - "bm25"
    - "hybrid"
    - "hybrid_rerank"

# Retry Configuration
retry:
  max_attempts: 3
  min_wait: 2  # seconds
  max_wait: 10  # seconds
  multiplier: 1

# Logging Configuration
logging:
  level: "INFO"
  format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
  file: "semantic_search.log"

# System Prompts
prompts:
  qa_system: |
    You're a helpful assistant.
    Please answer the following question {question} only using the following information {document}.
    If you can't answer the question, just say you can't answer that question.

  # Follow-up question prompt template
  follow_up_system: |
    You're a helpful assistant engaged in a conversation about a document.
    Use the previous conversation context to understand references and maintain continuity.

    Previous conversation:
    {conversation_context}

    Document context:
    {document}

    Please answer the following question: {question}
    If you can't answer the question, just say you can't answer that question.