config.yaml

# Document processing
chunk_size: 1000
overlap: 200

# File paths
data_dir: data/documents
output_dir: data/embeddings

# Logging
log_level: INFO

# Phase 3: cross-encoder reranking
reranker:
  model: cross-encoder/ms-marco-MiniLM-L-6-v2
  batch_size: 32
  score_threshold: 0.1
  top_k: 5

# Phase 3: context window (token counting for LLM prompt)
context:
  max_tokens: 4000
  tokenizer: gpt2

# Phase 3: RAG generation (Ollama)
generation:
  model: qwen2.5-coder:14b
  stream: false
  cache_ttl: 300

# Phase 4: multi-provider LLM routing
llm:
  default_provider: ollama
  default_model_by_provider:
    ollama: qwen2.5:7b
    openai: gpt-4o-mini
    anthropic: claude-sonnet-4-6
    gemini: gemini-2.5-flash
  allowed_models_by_provider:
    ollama:
      - qwen2.5:7b
      - deepseek-r1:8b
      - qwen2.5-coder:14b
    openai:
      - gpt-4o-mini
      - gpt-4.1-mini
    anthropic:
      - claude-sonnet-4-6
      - claude-haiku-4-5
    gemini:
      - gemini-2.5-flash
      - gemini-2.5-pro
  request_timeout_seconds: 60

# Phase 4: API production controls
api:
  auth_enabled: true
  api_keys:
    - dev-key-1
  rate_limit_per_minute: 120
  redis_rate_limit_enabled: true
  redis_url: redis://localhost:6379/0

# Evaluation settings
# Inline scoring adds ~200 ms per response (NLI model warm after first call).
# Set DOC_PROFILE=demo to use a lighter profile without inline NLI scoring.
evaluation:
  inline_enabled: true
  nli_model: cross-encoder/nli-deberta-v3-small