# RAG default configuration.
# Validates against config/schemas/rag-config.schema.json.
# See docs/06-ai-agents-rag/rag-architecture.md and rag-ingestion-and-chunking.md.

vectorStore:
  provider: pgvector                  # configurable: pgvector | azure_ai_search | qdrant | pinecone | weaviate
  namespaceStrategy: shared_index_with_filter   # or per_tenant_index for high-isolation tenants

chunking:
  strategy: semantic_hierarchical
  minTokens: 400
  maxTokens: 800
  overlapPercent: 12.5                # starting point per docs; tune via evaluation

ingestion:
  allowedSourceTypes:
    - policy
    - process
    - jd_reference
  excludeByDefault:
    - candidate_identity_documents
    - confidential_interview_feedback
  requiredMetadataFields:
    - tenantId
    - documentType
    - version
    - effectiveDate
    - expiryDate
    - accessPolicy
    - classification
    - source
    - headingPath
    - pageOrClauseRef
    - embeddingVersion
    - contentHash

retrieval:
  hybridSearchEnabled: true
  rerankingEnabled: true
  minRelevanceThreshold: 0.65
  maxContextTokens: 4000
  aclFilterBeforeSearch: true          # must remain true — see docs/adr/ADR-003
  noAnswerFallbackEnabled: true        # must remain true

evaluation:
  citationValidityMinPercent: 100
  groundednessMinScore: 0.85