# Semantic Retrieval Schema
# Based on REF-008 RAG Research
# Issues: #142, #143, #144

$schema: "https://json-schema.org/draft/2020-12/schema"
$id: "https://aiwg.io/schemas/semantic-retrieval/v1"
title: "Semantic Retrieval Configuration Schema"
description: |
  Schema for semantic retrieval, hybrid search, and context budget
  management per REF-008 RAG research.

type: object
required:
  - version
  - embedding
  - retrieval

properties:
  version:
    type: string
    pattern: "^\\d+\\.\\d+\\.\\d+$"
    default: "1.0.0"

  embedding:
    $ref: "#/$defs/EmbeddingConfig"

  retrieval:
    $ref: "#/$defs/RetrievalConfig"

  context_budget:
    $ref: "#/$defs/ContextBudgetConfig"

  hybrid_search:
    $ref: "#/$defs/HybridSearchConfig"

$defs:
  EmbeddingConfig:
    type: object
    description: "Embedding model configuration"
    properties:
      model:
        type: string
        default: "all-MiniLM-L6-v2"
        description: "Embedding model identifier"

      dimensions:
        type: integer
        default: 384
        description: "Embedding vector dimensions"

      cache_path:
        type: string
        default: ".aiwg/.cache/embeddings/"
        description: "Where to store cached embeddings"

      index_on_write:
        type: boolean
        default: true
        description: "Index artifacts when written"

      batch_size:
        type: integer
        default: 32
        description: "Batch size for embedding generation"

      indexed_paths:
        type: array
        items:
          type: string
        default:
          - ".aiwg/requirements/"
          - ".aiwg/architecture/"
          - ".aiwg/research/"
          - "src/"
          - "docs/"
        description: "Paths to index for retrieval"

      excluded_patterns:
        type: array
        items:
          type: string
        default:
          - "*.lock"
          - "node_modules/**"
          - ".git/**"
          - "*.min.js"
        description: "Patterns to exclude from indexing"

  RetrievalConfig:
    type: object
    description: "Retrieval settings"
    properties:
      top_k:
        type: integer
        default: 10
        description: "Number of results to return"

      similarity_threshold:
        type: number
        minimum: 0
        maximum: 1
        default: 0.7
        description: "Minimum similarity score to include"

      max_retrieval_time_ms:
        type: integer
        default: 2000
        description: "Maximum time for retrieval (ms)"

      rerank:
        type: boolean
        default: true
        description: "Apply reranking to results"

      rerank_model:
        type: string
        default: "cross-encoder/ms-marco-MiniLM-L-6-v2"
        description: "Model for reranking"

      prioritize_explicit_mentions:
        type: boolean
        default: true
        description: "Prioritize explicitly @-mentioned files"

      recency_boost:
        type: number
        minimum: 0
        maximum: 2
        default: 1.2
        description: "Boost factor for recently modified files"

  ContextBudgetConfig:
    type: object
    description: "Token budget management per REF-008"
    properties:
      total_budget_tokens:
        type: integer
        default: 200000
        description: "Total available context tokens"

      allocation:
        type: object
        properties:
          system_prompt:
            type: number
            default: 0.05
            description: "Fraction for system prompt"
          retrieved_context:
            type: number
            default: 0.65
            description: "Fraction for retrieved artifacts"
          generation:
            type: number
            default: 0.30
            description: "Fraction reserved for generation"
        description: "Budget allocation ratios (must sum to 1.0)"

      prioritization:
        type: object
        properties:
          explicit_mentions:
            type: integer
            default: 100
            description: "Priority score for explicit @-mentions"
          high_similarity:
            type: integer
            default: 80
            description: "Priority for high similarity (>0.9)"
          recent_modification:
            type: integer
            default: 60
            description: "Priority for recently modified (<24h)"
          same_directory:
            type: integer
            default: 40
            description: "Priority for same directory as task"
        description: "Priority scoring weights"

      overflow_strategy:
        type: string
        enum:
          - truncate_oldest
          - summarize_low_priority
          - exclude_low_similarity
          - hybrid
        default: hybrid
        description: "Strategy when budget exceeded"

      summarization:
        type: object
        properties:
          enabled:
            type: boolean
            default: true
          threshold_tokens:
            type: integer
            default: 5000
            description: "Summarize if artifact exceeds this"
          summary_target_tokens:
            type: integer
            default: 500
            description: "Target size for summaries"

      warnings:
        type: object
        properties:
          budget_warning_threshold:
            type: number
            default: 0.9
            description: "Warn when this fraction used"
          log_truncations:
            type: boolean
            default: true

  HybridSearchConfig:
    type: object
    description: "Two-stage hybrid retrieval per REF-008"
    properties:
      enabled:
        type: boolean
        default: true

      stage1_keyword:
        type: object
        properties:
          algorithm:
            type: string
            enum: [bm25, ripgrep, combined]
            default: ripgrep
          candidate_count:
            type: integer
            default: 100
            description: "Candidates from keyword stage"
          max_time_ms:
            type: integer
            default: 500
            description: "Max time for keyword search"

      stage2_semantic:
        type: object
        properties:
          rerank_count:
            type: integer
            default: 10
            description: "Final results after reranking"
          max_time_ms:
            type: integer
            default: 1500
            description: "Max time for semantic rerank"

      fusion:
        type: object
        properties:
          method:
            type: string
            enum: [rrf, weighted_sum, max_score]
            default: rrf
            description: "Score fusion method (Reciprocal Rank Fusion)"
          keyword_weight:
            type: number
            default: 0.4
          semantic_weight:
            type: number
            default: 0.6

# Agent protocol for retrieval
agent_protocol:
  auto_retrieval:
    description: "Automatic context retrieval when agent starts"
    triggers:
      - agent_task_start
      - context_refresh_request
    steps:
      - extract_task_query
      - run_hybrid_search
      - apply_context_budget
      - inject_into_context
      - log_retrieval_stats

  embedding_maintenance:
    description: "Keep embeddings current"
    triggers:
      - artifact_write
      - artifact_delete
      - manual_reindex
    steps:
      - detect_changed_files
      - generate_embeddings
      - update_index
      - prune_stale_entries

  budget_enforcement:
    description: "Enforce token budget"
    steps:
      - calculate_current_usage
      - check_against_budget
      - apply_overflow_strategy_if_needed
      - log_budget_state

# Performance targets
performance_targets:
  embedding_generation:
    per_file_ms: 100
    batch_throughput: "10 files/second"

  keyword_search:
    latency_p50_ms: 100
    latency_p99_ms: 500
    corpus_size: "10000 files"

  semantic_search:
    latency_p50_ms: 500
    latency_p99_ms: 2000
    corpus_size: "10000 files"

  combined_pipeline:
    total_latency_ms: 2000
    accuracy_target: "85% relevant in top 10"

# Telemetry
telemetry:
  metrics:
    - retrieval_latency_ms
    - keyword_candidates_count
    - final_results_count
    - average_similarity_score
    - budget_utilization_percent
    - truncation_frequency
    - cache_hit_rate

  logging:
    - query_text
    - retrieved_paths
    - similarity_scores
    - budget_state
    - overflow_actions

# Default configuration
defaults:
  storage_path: ".aiwg/.cache/retrieval/"
  index_file: "embeddings.index"
  metadata_file: "embeddings.meta.json"

# References
references:
  research:
    - "@.aiwg/research/findings/REF-008-rag.md"
  implementation:
    - "#142"  # Auto-retrieve artifacts
    - "#143"  # Hybrid retrieval
    - "#144"  # Token budget
  related:
    - "@src/core/context-manager.ts"
    - "@src/core/retrieval-pipeline.ts"