trevon/yams
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 2
yams/CHANGELOG.md
trvon df3997545b
 search benchmarking fixes
2026-01-19 02:50:05 -05:00

6.6 KiB
Raw Permalink Blame History

Changelog

All notable changes to YAMS (Yet Another Memory System) will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Archived Changelogs

  • v0.7.x archive: docs/changelogs/v0.7.md
  • v0.6.x archive: docs/changelogs/v0.6.md
  • v0.5.x archive: docs/changelogs/v0.5.md
  • v0.4.x archive: docs/changelogs/v0.4.md
  • v0.3.x archive: docs/changelogs/v0.3.md
  • v0.2.x archive: docs/changelogs/v0.2.md
  • v0.1.x archive: docs/changelogs/v0.1.md

[v0.7.11] - Unreleased

Breaking

  • Vector database migration required: sqlite-vec-cpp HNSW rewrite invalidates existing vector indices. After upgrading, run: 
    yams doctor repair --embeddings   # Regenerate all embeddings
yams doctor repair --graph        # Rebuild knowledge graph (optional)
    Without this, search will fall back to FTS5-only (no semantic search).
  • sqlite-vec-cpp submodule: HNSW API changes and third-party library removal (soft deletion, multi-threading, fp16 quantization, incremental persistence, pre-filtering).

Added

  • MCP graph tool for knowledge graph queries (parity with CLI yams graph).
  • Graph: snapshot-scoped version nodes, contains edges for file→symbol, --dead-code-report.
  • Graph prune policy (daemon.graph_prune) to keep latest snapshot versions.
  • Download CLI: progress streaming (human/json) via DownloadService callbacks.
  • Symbol-aware search ranking: definitions rank higher than usages (YAMS_SYMBOL_WEIGHT).
  • Zig language support: functions, structs, enums, unions, fields, imports, calls.
  • WEIGHTED_MAX fusion strategy: Takes maximum weighted score per document instead of sum. Prevents "hub" documents from dominating via multi-component consensus boost. Used by SCIENTIFIC tuning profile for benchmark corpora.

Performance

IPC & Daemon

  • IPC latency reduced from ~8-28ms to ~2-5ms (connection pooling, async timers, cached config).
  • Daemon startup throttling: PathTreeRepair via RepairCoordinator, Fts5Job startup delay (2s), reduced batch sizes (1000→100).

Ingestion & Storage

  • Post-ingest throughput: dedicated worker pool, adaptive backoff, batched directory ingests.
  • In-memory chunking for storeBytes() - avoids temp file I/O for large documents.

Database & Metadata

  • KGWriteQueue: Batched, serialized writes to KnowledgeGraphStore via async writer coroutine. Eliminates "database is locked" errors during high-throughput ingestion by queueing KG operations (nodes, edges, aliases, doc entities) and committing in batches. Both symbol extraction and NL entity extraction now use deferred batching with nodeKey→nodeId resolution at commit time.
  • Prepared statement caching for SQLite queries - reduces SQL compilation overhead on repeated operations. Cached methods: setMetadata, setMetadataBatch, getMetadata, getAllMetadata, getContent, getDocument, getDocumentByHash, updateDocument, deleteDocument, insertContent.
  • setMetadataBatch() API for bulk metadata updates - 4x faster than individual calls.

Search & Retrieval

  • Batch vector/KG lookups, flat_map for cache-friendly access, branch hints, memory pre-allocation.

Throughput Benchmarks (Debug, macOS M3 Max)

Benchmark Oct 2025 Jan 2026 Change
Ingestion_SmallDocument 2,771 ops/s 2,821 ops/s ~same
Ingestion_MediumDocument 56 ops/s 57 ops/s ~same
Ingestion_E2E (100 docs) - 9.2 docs/s new (KGWriteQueue)
Metadata_SingleUpdate 10,537 ops/s 17,794 ops/s +69%
Metadata_BulkUpdate(500) 7,823 ops/s 50,473 ops/s +6.5x
IPC_StreamingFramer - 3,732 ops/s new
IPC_UnaryFramer - 10,088 ops/s new

Experimental

  • libSQL backend: Default database backend with concurrent write support via MVCC. Enables up to 4x write throughput during heavy indexing. Configure with meson option database-backend (choices: libsql [default], sqlite).

    Installation: If Rust toolchain is available, libsql builds automatically from source via the meson subproject. Otherwise falls back to SQLite.

    # Ensure Rust is installed (for automatic build)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
# Or disable libsql to use standard SQLite
meson configure -Ddatabase-backend=sqlite

    See libSQL for details.

Documentation

  • Embedding model recommendations: Added model comparison table to README. 384-dim models (e.g., all-MiniLM-L6-v2) recommended for best speed/quality tradeoff.

Changed

  • Reranking: Score-based reranking is now the default. Uses geometric mean of text and vector scores to boost documents with multi-component consensus. No external model needed. Cross-encoder model reranking is opt-in via enableModelReranking config option.

Fixed

  • FTS5 natural language queries: OR fallback now correctly triggers when AND query returns zero results. Previously, long queries like scientific abstracts would fail because the AND query returned nothing and the OR fallback condition was never met.
  • ONNX multi-threading on Linux/macOS: Removed forced single-threaded execution that was only needed for Windows. Non-Windows platforms now use intra_op_threads=4 by default, improving inference speed for 768-dim and larger models by 2-4x.
  • Hybrid search fusion: fallback to non-empty filePath when vector results have empty paths (hash→path lookup failures no longer cause result mismatches).
  • TSAN race in daemon_search(): pass DaemonSearchOptions by value to avoid stack reference escaping to coroutine thread.
  • TSAN race in handle_streaming_request(): check connection_closing_ before socket.is_open() to avoid race with handle_connection closing the socket.
  • Compression stats now persist across daemon restarts (Storage Logical Bytes vs CAS Unique Raw Bytes now show correct values).
  • CLI rejects ambiguous subcommands (e.g., yams search graph → use --query).
  • --paths-only search now returns results correctly.
  • yams watch waits for daemon ready; always ignores .git contents.
  • Expanded prune patterns for build artifacts and language caches.
  • Fixed ONNX model loading deadlock on Windows (single-flight pattern, recursive mutex).
  • Streaming: 30s chunk timeout, backpressure stops producer on queue overflow.
  • yams add returns immediately; hash computed async during ingestion.
  • Replaced experimental Boost.Asio APIs with stable async_initiate (fixes TSAN races).