# Changelog All notable changes to the Vector Bridge MCP Server will be documented in this file. The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [1.3.0] - 2025-10-06 ### Added - Phase 4: Smart Solution Discovery & Pattern Surfacing - **Pattern-Solution Linking**: Automatic linking of detected patterns to proven solutions - **Pattern Detection in Queries**: `detect_patterns_in_query()` analyzes error messages to find matching patterns - **Pattern-Based Solution Ranking**: Solutions ranked by pattern-specific success rates - **Golden Paths**: Discover most successful pattern-solution combinations across projects - **Cross-Project Pattern Intelligence**: Learn which solutions work best for specific error patterns - **Migration 007**: Pattern-solutions table and smart discovery functions ### New MCP Tools - `pattern_detect`: Detect patterns in error messages and suggest linked solutions - `pattern_solutions`: Get solutions ranked for a specific pattern (e.g., "redis-connection") - `pattern_link`: Link a pattern to a solution and record success/failure for learning - `golden_paths`: Get proven pattern-solution combinations with high success rates ### SolutionProvider API - `detectPatternsInQuery()`: Analyze query text for patterns - `getSolutionsForPattern()`: Get ranked solutions for a pattern tag - `linkPatternToSolution()`: Record pattern-solution application with outcome - `getGoldenPaths()`: Retrieve most successful pattern-solution pairs - `updatePatternSolutionHelpfulness()`: Sync feedback scores to pattern rankings ### Changed - Solutions now track pattern-specific success rates in addition to global success rates - Pattern-solution ranking combines: success rate (60%) + applications (30%) + helpfulness (10%) - Golden paths require 3+ applications by default to ensure statistical significance ### Technical Details - Pattern tags derived from `documents.tags` field (first tag used as pattern identifier) - Pattern categories align with solution categories (runtime, build, deploy, etc.) - Pattern detection requires minimum 3 document occurrences to avoid noise - Pattern matching uses tag overlap (minimum 10% match score required) - Success tracking: `success_count` and `failure_count` per pattern-solution pair ## [1.2.0] - 2025-10-06 ### Added - Phase 3: Learning System - **Memory Feedback Tool** (`memory_feedback`): Record helpful/unhelpful feedback on search results - **Feedback Bonus**: +15% ranking boost for memories with consistently helpful feedback - **Pattern Detection**: SQL functions to detect recurring solutions across projects (`detect_patterns`) - **Memory Feedback Table**: Tracks helpful/unhelpful ratings with context - **Enhanced Search Functions**: `search_hybrid_with_feedback` and `search_hybrid_global_with_feedback` - **Migration 006**: Feedback system and pattern detection ### Changed - Search results now include `chunk_id` in meta for feedback tracking - Search results include `feedback_score` showing helpful ratio (0-1) - PgVectorProvider now uses feedback-enhanced search functions - Search rankings self-improve over time based on actual usefulness ### Fixed - Integrated feedback bonus into actual search queries (was calculated but not used) - Cache keys now differentiate between feedback-enhanced and standard search ## [1.1.0] - 2025-10-06 ### Added - Phase 2: Hybrid Search - **Hybrid Search Ranking**: Combines vector similarity (60%) + BM25 text search (30%) + time decay (10%) - **BM25 Full-Text Search**: PostgreSQL `tsvector` and `ts_rank_cd` for keyword matching - **Time Decay**: Exponential decay with 30-day half-life favors recent memories - **Outcome Bonus**: +10% for successful solutions, -5% for failures - **Conversation Summary Ingestion**: PostCompact hook (`conversation_summary_ingest.py`) - **Auto Context Injection**: PreToolUse hook (`memory_context_inject.py`) - **Migration 005**: Hybrid search functions and `chunk_tsv` column ### Changed - Search results now include detailed scoring breakdown: `vector_score`, `bm25_score`, `time_score`, `outcome_bonus` - Cache parameters include `hybrid: true` flag to differentiate from old vector-only cache - Query text now used for BM25 matching in addition to vector similarity ### Performance - BM25 index (`documents_chunk_tsv_idx`) for fast full-text search - Automatic `tsvector` updates via trigger - Redis query cache (5-minute TTL, 14.6x speedup) ## [1.0.0] - 2024-10-04 ### Added - Foundation - Multi-tenant vector memory with project isolation - OpenAI embeddings (`text-embedding-3-small`, 1536 dimensions) - Postgres + pgvector for vector storage - Redis cache for embeddings (60-day TTL, ~70% cost savings) - Basic vector search with cosine similarity - Automatic text chunking (500-800 tokens with overlap) - Dedupe system (SHA256 content hashing, 48-hour cache) - Solution memory system for error fixpacks ### MCP Tools - `memory_ingest`: Ingest text content with metadata - `memory_search`: Semantic search with filters - `memory_projects`: List indexed projects - `solution_search`: Find fixpacks for error messages - `solution_apply`: Record success/failure of solutions - `solution_preview`: DRY-RUN preview of solution steps - `solution_upsert`: Create/update solution fixpacks ### Migrations - `001_init.sql`: Base schema (projects, documents, search functions) - `002_add_metadata.sql`: Categorization (component, category, tags) - `003_solution_memory.sql`: Solution fixpacks - `004_hnsw_indexes.sql`: HNSW indexes for performance ### Hook Integration - `stop_digest.py`: Automatic DIGEST ingestion after each session - Queue-based async ingestion with retry logic ### Performance - HNSW indexes for fast approximate nearest neighbor search - Redis embedding cache (60-day TTL) - Dedupe cache (48-hour TTL) - Graceful degradation when Redis unavailable --- ## Future Releases ### [1.3.0] - Planned - A/B testing framework for ranking algorithms - Temporal reasoning ("what worked last month?") - Multi-modal memory (code + diagrams + screenshots) - Personalized rankings per user - Cross-project pattern catalog --- ## Migration Notes ### Upgrading from 1.1.0 to 1.2.0 Run migration 006: ```bash psql $DATABASE_URL_MEMORY < migrations/006_feedback_system.sql ``` ### Upgrading from 1.0.0 to 1.1.0 Run migration 005: ```bash psql $DATABASE_URL_MEMORY < migrations/005_hybrid_search.sql ``` --- ## Breaking Changes ### None in 1.2.0 - All changes are backward compatible - Existing search queries work without modification - Feedback system is opt-in ### None in 1.1.0 - All changes are backward compatible - Old `search_documents` functions still work - Hybrid search is used automatically for better results --- ## Contributors - Main Agent (Claude Orchestrator) - Implementation Engineer - Database Modeler - Test Author Generated with [Claude Code](https://claude.com/claude-code)