config.yaml

## AutoRAG Configuration
## This file defines all tunable parameters across the 7 optimization dimensions.
## The optimizer agent modifies this file between experiments.
## Changes to chunking or embedding sections require re-indexing the vector store.

# Dimension 1: Chunking Strategy
chunking:
  strategy: "fixed"          # fixed | sentence | paragraph
  chunk_size: 512            # 256 | 512 | 1024 | 2048
  chunk_overlap: 100         # 0 | 50 | 100 | 200

# Dimension 2: Embedding Model
embedding:
  provider: "openai"         # voyage | openai
  model: "text-embedding-3-small"   # voyage-3-large | voyage-3.5-lite | text-embedding-3-small
  dimensions: 1536           # 1024 (voyage-3-large) | 512 (voyage-3.5-lite) | 1536 (text-embedding-3-small)

# Dimension 3: Retrieval Parameters
retrieval:
  top_k: 5                  # 3 | 5 | 8 | 10 | 15
  search_type: "vector"     # vector | fts | hybrid
  reranking: false           # true | false
  distance_metric: "cosine"  # cosine | l2 | dot

# Dimension 4: Model Routing
models:
  query_classifier:
    model: "claude-haiku-4-5-20251001"
    temperature: 0.0
    max_tokens: 256
  query_rewriter:
    model: "claude-haiku-4-5-20251001"
    temperature: 0.0
    max_tokens: 512
  answer_generator:
    model: "claude-sonnet-4-6"
    temperature: 0.0
    max_tokens: 1024
  answer_validator:
    model: "claude-sonnet-4-6"
    temperature: 0.0
    max_tokens: 256

# Dimension 5: Few-Shot Example Selection
few_shot:
  enabled: false             # start without, optimizer can enable
  strategy: "none"           # none | fixed | domain_matched | difficulty_matched
  examples_per_query: 0      # 0 | 1 | 2 | 3
  example_pool: "curated"    # curated | random_sample

# Dimension 6: Pipeline Topology
pipeline:
  query_classification: true
  query_rewriting: true
  multi_step_retrieval: false
  answer_validation: true
  confidence_threshold: 0.80  # below this, return "I don't know"
  false_premise_detection: true