agentic-flow 2.0.1-alpha.17 → 2.0.1-alpha.19

package/CHANGELOG.md

@@ -2,6 +2,82 @@
 
 All notable changes to this project will be documented in this file.
 
+## [2.0.1-alpha.19] - 2025-12-31
+
+### Added
+- **Cache Pretrain**: Pre-populate embedding cache for faster cold starts
+  - `pretrain(sources)` - Pretrain from files, globs, or text arrays
+  - `pretrainCodePatterns()` - 45 common programming patterns (~3ms)
+  - `pretrainFromRepo(path)` - Crawl and cache entire repositories
+
+- **AI-Enhanced Pretrain** (ruvector integration):
+  - `pretrainWithAI()` - Uses attention mechanisms for intelligent caching
+  - `pretrainIntelligent()` - 5-stage analysis (patterns, AST, git, deps, chunks)
+  - `pretrainIncremental()` - Only cache changed files (git diff)
+  - `pretrainSemantic()` - Code-aware chunking (functions, classes, interfaces)
+  - `prefetchForContext()` - Context-aware prefetch based on current task
+
+- **Attention-Guided Caching**:
+  - `HyperbolicAttention` - Hierarchical code structure (AST trees)
+  - `MoEAttention` - Expert routing (frontend/backend/data/test)
+  - `GraphRoPeAttention` - Dependency graph understanding
+  - `FastGRNN` - Pattern prediction for smart prefetch
+
+- **Smart Pretraining Features**:
+  - Semantic chunking by code boundaries (not fixed size)
+  - Priority-based caching of frequently accessed patterns
+  - Background pretraining (non-blocking)
+  - Git history analysis (cache hot files first)
+  - Dependency graph caching
+
+### Performance (pretrain)
+| Operation | Time | Notes |
+|-----------|------|-------|
+| Code patterns | ~3ms | 45 patterns, one-time |
+| Custom texts | ~0.3ms | 5 texts batch |
+| Cached query | ~0.01ms | Memory cache hit |
+| Incremental | ~50ms | Git-based, 20 commits |
+| AI pretrain | ~200ms | Full 5-stage analysis |
+| Context prefetch | ~5ms | Task-aware prediction |
+
+## [2.0.1-alpha.18] - 2025-12-31
+
+### Added
+- **Persistent SQLite Embedding Cache**: Cross-session embedding persistence
+  - New `EmbeddingCache` class with SQLite (better-sqlite3) backend
+  - **9000x faster** cached lookups: ~0.045ms vs ~400ms ONNX inference
+  - Batch writes with transactions: 0.016ms per embedding
+  - Stored at `~/.agentic-flow/embedding-cache.db`
+  - LRU eviction with configurable max entries (default: 10,000)
+  - Automatic cleanup of entries older than 30 days
+  - WAL mode for concurrent access
+
+- **Cache Hierarchy** (checked in order):
+  1. In-memory LRU cache (~0.001ms) - fastest, ephemeral
+  2. Persistent SQLite cache (~0.045ms) - persists across sessions
+  3. ONNX inference (~400ms) - only on complete cache miss
+
+- **EmbeddingService Cache Methods**:
+  - `clearPersistentCache()` - Clear SQLite cache
+  - `clearAllCaches()` - Clear both memory and persistent caches
+  - `getPersistentCacheStats()` - Get cache statistics
+
+- **Environment Variables**:
+  - `AGENTIC_FLOW_PERSISTENT_CACHE=true|false` (default: true)
+
+### Performance (with persistent cache)
+| Operation | Time | Notes |
+|-----------|------|-------|
+| ONNX cold | ~400ms | First inference + cache write |
+| SQLite read | ~0.045ms | Persistent cache lookup |
+| SQLite batch write | 0.016ms/ea | Transactional batch insert |
+| Memory cache | ~0.001ms | In-memory LRU lookup |
+| **Speedup** | **9000x** | SQLite vs ONNX inference |
+
+### Changed
+- EmbeddingService now checks persistent cache before ONNX inference
+- Only semantic (ONNX) embeddings are persisted (simple hash-based are not)
+
 ## [2.0.1-alpha.17] - 2025-12-31
 
 ### Added