config.yml

# =============================================================================
# CookHero Configuration File
# =============================================================================
# This file contains all application configuration for the CookHero RAG system.
# Sensitive values (API keys, passwords) should be stored in .env file.
#
# Configuration is organized into logical sections:
# 1. LLM Provider Configuration (layered: fast / normal)
# 2. Embedding Model Configuration
# 3. Vector Store Configuration
# 4. RAG Pipeline Configuration (retrieval, reranking, cache)
# 5. Vision/Multimodal Configuration
# 6. Evaluation Configuration
# 7. Web Search Configuration
# 8. Data Source Configuration
# 9. Database Connection Configuration
# =============================================================================

# =============================================================================
# 1. LLM Provider Configuration (Layered: fast / normal)
# =============================================================================
# The system uses two tiers of LLM models:
#   - fast: Optimized for low latency (intent detection, query rewriting)
#   - normal: Optimized for quality (final response generation)
#
# Each tier supports:
#   - model_names: List of models (one randomly selected per call)
#   - base_url: API endpoint URL
#   - api_key: Loaded from .env (FAST_LLM_API_KEY / LLM_API_KEY)
#   - temperature: Randomness control (0.0-2.0)
#   - max_tokens: Maximum tokens to generate
# =============================================================================

llm:
  # Fast tier - Low latency models for auxiliary tasks
  fast:
    model_names:
      - "XiaomiMiMo/MiMo-V2-Flash"
      # - "deepseek-ai/DeepSeek-V3.2"
      - "Qwen/Qwen3-235B-A22B-Instruct-2507"
      - "Qwen/Qwen3-VL-235B-A22B-Instruct"
    base_url: "https://api-inference.modelscope.cn/v1"
    # api_key: Loaded from .env (FAST_LLM_API_KEY or LLM_API_KEY)
    temperature: 1
    max_tokens: 131072

  # Normal tier - High quality models for main generation
  normal:
    model_names:
      - "Qwen/Qwen3-VL-30B-A3B-Instruct"
      - "deepseek-ai/DeepSeek-V3.2"
    base_url: "https://api.siliconflow.cn/v1"
    # api_key: Loaded from .env (LLM_API_KEY)
    temperature: 1
    max_tokens: 131072

# =============================================================================
# 2. Embedding Model Configuration
# =============================================================================
# Used for converting text to vectors for semantic search.
# =============================================================================

embedding:
  # HuggingFace model name for text embedding
  # BGE models are optimized for Chinese text
  model_name: "BAAI/bge-small-zh-v1.5"

# =============================================================================
# 3. Vector Store Configuration
# =============================================================================
# Milvus vector database configuration.
# Connection details are in the database.milvus section below.
# =============================================================================

vector_store:
  # Currently only Milvus is supported
  type: "milvus"

  # Collection names for different data types
  collection_names:
    # Global recipe collection (HowToCook)
    recipes: "cook_hero_recipes"
    # User's personal documents collection
    personal: "cook_hero_personal_docs"

# =============================================================================
# 4. RAG Pipeline Configuration
# =============================================================================

# 4.1 Retrieval Configuration
# ------------------------------------------------------------------------------
retrieval:
  # Number of documents to retrieve
  top_k: 9
  # Score threshold for filtering results (0.0-1.0)
  score_threshold: 0.2
  # Ranking method: "weighted" or "rrf" (Reciprocal Rank Fusion)
  ranker_type: "weighted"
  # Weights for hybrid ranking [dense_vector, sparse_bm25]
  ranker_weights: [0.8, 0.2]

# 4.2 Reranker Configuration
# ------------------------------------------------------------------------------
# Reranker further refines retrieval results for better relevance.
# Uses the normal LLM API key by default (can override with RERANKER_API_KEY)
reranker:
  enabled: true
  # Reranker provider type
  type: "siliconflow"
  # Model name for reranking
  model_name: "Qwen/Qwen3-Reranker-8B"
  # API endpoint for reranking
  base_url: "https://api.siliconflow.cn/v1/rerank"
  # api_key: Loaded from .env (RERANKER_API_KEY or falls back to LLM_API_KEY)
  temperature: 0.0
  max_tokens: 8192
  # Minimum score threshold for reranked results
  score_threshold: 0.2

# 4.3 Cache Configuration
# ------------------------------------------------------------------------------
# Two-level caching strategy:
#   - L1: Exact match (Redis) - fast lookup for identical queries
#   - L2: Semantic match (Milvus) - handles similar queries via vector similarity
#
# Note: Only caches Query -> Retrieved Documents (Context), NOT LLM responses.
# Connection settings are in database.redis and database.milvus sections below.
# ------------------------------------------------------------------------------
cache:
  enabled: true
  # Cache TTL in seconds (applies to both L1 and L2)
  ttl: 3600  # 1 hour

  # L2 Semantic Cache Configuration
  l2_enabled: true
  # Similarity threshold for L2 cache hits (0.0-1.0)
  similarity_threshold: 0.92
  # Collection name for L2 cache storage
  vector_collection: "cookhero_retrieval_cache"
  # L2 cache credentials loaded from .env (MILVUS_USER / MILVUS_PASSWORD)

# =============================================================================
# 5. Vision/Multimodal Configuration
# =============================================================================
# Configuration for image analysis and multimodal capabilities.
# =============================================================================

vision:
  model:
    # Enable/disable vision features
    enabled: true
    # Model configuration (OpenAI-compatible API)
    model_name: "Qwen/QVQ-72B-Preview"
    base_url: "https://api-inference.modelscope.cn/v1"
    # api_key: Loaded from .env (VISION_API_KEY or LLM_API_KEY)
    temperature: 0.7
    max_tokens: 4096
    # Image processing settings
    max_image_size_mb: 10.0
    # Maximum request timeout in seconds
    request_timeout: 120

# =============================================================================
# 6. RAG Evaluation Configuration (RAGAS)
# =============================================================================
# Configuration for quality monitoring using the RAGAS framework.
# =============================================================================

evaluation:
  enabled: true
  # Async mode: Run evaluation in background without blocking response
  async_mode: true
  # Sampling rate (0.0-1.0): Fraction of requests to evaluate
  # 1.0 = evaluate all requests, 0.1 = evaluate 10% of requests
  sample_rate: 1.0
  # RAGAS metrics to compute
  # Available: "faithfulness", "answer_relevancy", "context_precision", "context_recall"
  metrics:
    - "faithfulness"
    - "answer_relevancy"
  # LLM tier to use for evaluation (fast recommended for cost efficiency)
  llm_type: "fast"
  # Evaluation timeout in seconds
  timeout_seconds: 600
  # Quality alert thresholds
  alert_thresholds:
    faithfulness: 0.3
    answer_relevancy: 0.5

# =============================================================================
# 7. Web Search Configuration (Tavily)
# =============================================================================
# Configuration for real-time web search enhancement.
# =============================================================================

web_search:
  enabled: true
  # Maximum number of search results to return
  max_results: 6
  # api_key: Loaded from .env (WEB_SEARCH_API_KEY)

# =============================================================================
# 8. Image Generation Configuration (OpenAI DALL-E)
# =============================================================================
# Configuration for AI image generation using OpenAI's DALL-E 3 model.
# =============================================================================

image_generation:
  enabled: true
  model: "Kwai-Kolors/Kolors"
  base_url: "https://api.siliconflow.cn/v1"
  # api_key: Loaded from .env (OPENAI_IMAGE_API_KEY)

# =============================================================================
# 9. MCP (Model Context Protocol) Configuration
# =============================================================================
# Configuration for MCP server integrations.
# =============================================================================

mcp:
  # Amap (高德地图) MCP server configuration
  amap:
    enabled: true
  # amap_api_key: Loaded from .env (AMAP_API_KEY)

# =============================================================================
# 10. Data Source Configuration
# =============================================================================
# Configuration for external data sources (e.g., HowToCook recipe library).
# =============================================================================

data_source:
  howtocook:
    # Subdirectory path for dishes
    path_suffix: "dishes"
    # Subdirectory path for cooking tips
    tips_path_suffix: "tips"
    # Markdown headers to split documents on
    headers_to_split_on:
      - ["#", "header_1"]
      - ["##", "header_2"]

# =============================================================================
# 11. Database Connection Configuration
# =============================================================================
# Connection details for PostgreSQL, Redis, and Milvus.
# Passwords should be loaded from .env file for security.
# =============================================================================

database:
  # PostgreSQL - Primary relational database
  postgres:
    host: "localhost"
    port: 5432
    database: "cookhero"
    user: "cookhero"
    # password: Loaded from .env (DATABASE_PASSWORD)
    # Connection pool settings
    pool_size: 20
    max_overflow: 30
    pool_timeout: 30
    pool_recycle: 1800
    # Echo SQL queries (for debugging)
    echo: false

  # Redis - Key-value store and L1 cache
  redis:
    host: "localhost"
    port: 6379
    db: 0
    # password: Loaded from .env (REDIS_PASSWORD, optional)

  # Milvus - Vector database for embeddings and L2 cache
  milvus:
    host: "localhost"
    port: 19530
    # user: Loaded from .env (MILVUS_USER, optional)
    # password: Loaded from .env (MILVUS_PASSWORD, optional)
    # secure: false  # Set to true for TLS connection

# =============================================================================
# 12. Path Configuration
# =============================================================================
# File system paths for data storage.
# =============================================================================

paths:
  # Base path for the HowToCook recipe data
  base_data_path: "data/HowToCook"

# =============================================================================
# 13. GitHub Issues Configuration (Optional)
# =============================================================================
# Configuration for the self-hosted GitHub issues feature.
# =============================================================================

blank_issues_enabled: true
contact_links: []