config.yaml.example

# =============================================================================
# Unified RAG configuration template for dynamic_retriever.py
#
# This template supports BOTH paradigms exposed by dynamic_retriever.py:
#   - tool-use     : agentic LLM-driven tool-calling loop (default)
#   - decomposed   : multi-round decomposed RAG (DecomposedRAGPipeline)
#
# Select via `pipeline.mode` below or `--mode {tool-use,decomposed}` on the CLI
# (CLI overrides config).
#
# Section comments mark each block as [shared], [tool-use only], or [decomposed only].
# =============================================================================

# [shared] Paths (relative to working directory)
paths:
  questions_path: ""
  output_root: ""

# [tool-use only] Max docs to keep after a `prune` tool call / auto-prune
prune_k: 20
# [tool-use only] Enable HyDE passage generation before embedding the query
hyde: false

# [shared] Default pipeline paradigm; CLI `--mode` overrides this value.
pipeline:
  mode: tool-use            # {tool-use, decomposed}
  # [decomposed only] Sub-question decomposition + rounds (DecomposedRAGPipeline)
  max_sub_questions: 5
  subq_fanout_cap: 3
  subq_max_concurrency: 2
  rounds: 2
  preliminary_prefix: ""
  subquery_prefix: ""
  dataset_description: ""

# [decomposed only] Retrieval parameters for diversity/log-det selection
retrieval:
  k_fulltext: 10
  k_diverse: 0              # 0 = disabled; >0 = select this many diverse chunks via log-det
  eta: 0.0
  rescale_power: 5.0
  cache_size: 2000

# [decomposed only] Concurrency / question-count limits for main_async()
execution:
  max_questions: null       # null = process all questions
  max_workers: 8

# Azure OpenAI / LLM settings ([shared]).
llm:
  llm_endpoint: "https://<foundry-resource-name>.cognitiveservices.azure.com/"
  api_version: "2024-12-01-preview"
  llm_model: "<deployment-name>"
  temperature: 0.0
  max_completion_tokens: 2048
  max_retries: 2
  premium_max_concurrency: 8
  prompt_cache_size: 4096
  use_rbac_auth: true
  token_scope: "https://cognitiveservices.azure.com/.default"
  llm_api_key: ""
  context_limit: 270000     # [tool-use only] explicit context window hint

# [decomposed only] Optional separate LLM endpoint for regeneration steps in efficient pipeline
llm_regen:
  use_llm_regen: false
  llm_endpoint: ""
  api_version: ""
  llm_model: ""
  temperature: 0.0
  max_completion_tokens: 4096
  max_retries: 5
  premium_max_concurrency: 4
  use_rbac_auth: false
  token_scope: ""
  llm_api_key: ""

# [decomposed only] Ollama (or other local) fallback for sub-tasks
local_llm:
  local_fallback_endpoint: "http://localhost:11434/api/generate"
  local_fallback_model: ""
  use_local_fallback_for_subtasks: false
  local_max_concurrency: 8

# Embedding settings ([shared])
embedding:
  embed_endpoint: "https://<foundry-resource-name>.openai.azure.com/"
  api_version: "2024-12-01-preview"
  embed_model: "<deployment-name>"
  embed_dimensions: 1024
  embed_cache_size: 4096
  use_rbac_auth: true
  token_scope: "https://ai.azure.com/.default"
  embed_api_key: ""


# Semantic ranker ([shared] for use_ranker/region/account_name/tenant_id/token_scope/batch_size/max_retries)
# - rerank_multiplier : [tool-use only] proportional reranking
# - k_ranker          : [decomposed only] absolute reranking (after diversity selection)
ranker:
  use_ranker: true
  rerank_multiplier: 8      # [tool-use only]
  k_ranker: 20              # [decomposed only]
  region: ""
  account_name: ""
  tenant_id: ""
  token_scope: ""
  batch_size: 16
  max_retries: 5

# Cosmos DB settings
cosmos:
  uri: ""
  key: ""
  database_name: ""
  use_rbac_auth: false

  # Upload settings
  cosmos_account_name: ""
  cosmos_resource_group: ""
  azure_subscription_id: ""
  embedding_batch_size: 20

  vector_embedding_policy_json: |
    {
      "vectorEmbeddings": [
        {
          "path": "/embedding",
          "dataType": "float32",
          "dimensions": 1024,
          "distanceFunction": "cosine"
        }
      ]
    }

  # Configure each source/container independently
  sources:
    - id: "source_1"
      container_name: "container_1"
      partition_key_path: "/pk"
      embedding_field: "embedding"
      documents_root: ""
      embedding_text_fields:
        - title
        - summary
        - content
      retrieval:
        search_k: 10
        fulltext_search_k: 10
        fulltext_fields:
          - title
      indexing_policy_json: |
        {
          "indexingMode": "consistent",
          "automatic": true,
          "includedPaths": [
            { "path": "/*" }
          ],
          "excludedPaths": [
            { "path": "/\"_etag\"/?" },
            { "path": "/embedding/*" }
          ],
          "fullTextIndexes": [
            { "path": "/title" }
          ],
          "vectorIndexes": [
            {
              "path": "/embedding",
              "type": "diskANN",
              "quantizationByteSize": 192,
              "indexingSearchListSize": 100
            }
          ]
        }
      full_text_policy_json: |
        {
          "defaultLanguage": "en-US",
          "fullTextPaths": [
            { "path": "/title", "language": "en-US" }
          ]
        }

    - id: "source_2"
      container_name: "container_2"
      partition_key_path: "/pk"
      embedding_field: "embedding"
      documents_root: ""
      embedding_text_fields:
        - text
      retrieval:
        search_k: 25
        fulltext_search_k: 0
        fulltext_fields: 
          - text
      indexing_policy_json: |
        {
          "indexingMode": "consistent",
          "automatic": true,
          "includedPaths": [
            { "path": "/*" }
          ],
          "excludedPaths": [
            { "path": "/\"_etag\"/?" },
            { "path": "/embedding/*" }
          ],
          "fullTextIndexes": [
            { "path": "/text" }
          ],
          "vectorIndexes": [
            {
              "path": "/embedding",
              "type": "diskANN",
              "quantizationByteSize": 192,
              "indexingSearchListSize": 100
            }
          ]
        }
      full_text_policy_json: |
        {
          "defaultLanguage": "en-US",
          "fullTextPaths": [
            { "path": "/text", "language": "en-US" }
          ]
        }