config/default_config.yaml

# Default configuration for the Hugging Face Spaces deployment (API-first).
# IMPORTANT: do NOT put any real API keys in this file. Use Spaces Secrets instead.

# Prebuilt index (vector DB + doc store) stored as a Hugging Face dataset repo.
index:
  repo_id: "${RAG_INDEX_REPO_ID:ZhangNy/radiology-index-qwen3-embedding-0.6b}"
  revision: "${RAG_INDEX_REVISION:main}"

# Embedding configuration (query-time embeddings must match the index embedding model).
embedding:
  type: "api"  # Options: "api" or "local" (local requires requirements-dev.txt)
  api_base_url: "${EMBED_API_BASE_URL:https://api.siliconflow.cn/v1}"
  api_key: "${EMBED_API_KEY:}"
  model_name: "${EMBED_MODEL_NAME:Qwen/Qwen3-Embedding-0.6B}"
  batch_size: 32

# Reranker configuration (recommended; can be disabled if no key is provided).
reranker:
  enabled: true
  type: "api"  # Options: "api" or "local" (local requires requirements-dev.txt)
  api_base_url: "${RERANK_API_BASE_URL:https://api.siliconflow.cn/v1}"
  api_key: "${RERANK_API_KEY:}"
  model_name: "${RERANK_MODEL_NAME:BAAI/bge-reranker-v2-m3}"
  top_k: 10

# LLM configuration (OpenAI-compatible API).
llm:
  base_url: "${LLM_BASE_URL:https://poloai.top/v1}"
  api_key: "${LLM_API_KEY:}"
  model_name: "${LLM_MODEL_NAME:gemini-3-flash-preview}"
  temperature: 0.7
  max_tokens: 2000

# Storage paths (prefer /data on Spaces; app.py will default RAG_STORAGE_DIR to /data/radiology_rag when available).
storage:
  vector_db_path: "${RAG_STORAGE_DIR:./storage}/chroma_db"
  doc_store_path: "${RAG_STORAGE_DIR:./storage}/doc_store.db"

# Text splitting parameters (used for index build scripts; kept here for transparency).
processing:
  chunk_size: 1024
  chunk_overlap: 200
  separators:
    - "\n\n#### "
    - "\n\n### "
    - "\n\n## "
    - "\n\n"
    - "\n"
    - " "
    - ""
  keep_separator: true

# Retrieval configuration
retrieval:
  # Default strategy for this Space:
  # - balanced_multi_source: includes Wikipedia (encyclopedia) by default
  strategy: "balanced_multi_source"
  top_k: 20
  source_filters:
    - "article"
    - "case"
    - "tutorial"
    - "encyclopedia"

  search_type: "similarity"  # "similarity" or "mmr"
  chunk_fetch_multiplier: 3

  # MMR parameters (only if search_type == "mmr")
  mmr_lambda: 0.5
  mmr_fetch_k: 50

  # Balanced multi-source retrieval policy
  multi_source:
    total_top_k: 8
    sources_priority: ["article", "case", "encyclopedia", "tutorial"]
    article:
      candidate_k: 80
      max_k: 3
      min_score: 0.15
      required: true
    case:
      candidate_k: 80
      max_k: 3
      min_score: 0.15
      required: true
    encyclopedia:
      candidate_k: 8
      max_k: 2
      min_score: 0.15
      required: true
    tutorial:
      candidate_k: 20
      max_k: 2
      min_score: 0.50
      required: false

# Encyclopedia configuration (Wikipedia)
encyclopedia:
  wikipedia:
    language: "en"
    user_agent: "RadiologyRAG-Space/1.0"
    timeout_s: 10
    max_chars_per_doc: 2000

# Citation configuration (no images in this Space; references link to original pages)
citation:
  format: "numbered"
  max_content_length: 900

# Gradio UI configuration
ui:
  title: "Thoracic Radiology RAG System"
  description: "Ask questions about thoracic radiology and get answers with citations (articles, cases, tutorials + Wikipedia)."
  theme: "soft"
  show_retrieved_docs: true

# Logging configuration
logging:
  level: "INFO"
  format: "%(asctime)s - %(levelname)s - %(message)s"