Skip to content

CHANGELOG.md

Latest commit

 

History

History
113 lines (80 loc) · 4.52 KB

CHANGELOG.md

File metadata and controls

113 lines (80 loc) · 4.52 KB

Changelog

All notable changes to RecallForge will be documented in this file.

[Unreleased]

Nothing yet.

[0.2.0] — 2026-03-22

RecallForge 0.2.0 makes the project much more usable as a local multimodal memory MCP. This release improves cross-modal retrieval, adds a stronger MCP surface for agents, and lays the foundation for memory-level results instead of isolated file fragments.

Note: Python 3.12 and 3.13 supported. Python 3.14 is not yet supported due to the pinned pyarrow wheel range for this release.

Memory MCP

  • First-class memory_id support in storage and search
  • Parent/child memory linkage for derived assets such as frames, transcripts, and document sections
  • Memory-centric MCP capabilities including memory lookup and listing
  • memory:// resources for stable memory access
  • Memory-level result rollup so related assets can surface as a single memory hit

Search and Retrieval

  • Hybrid search support for image and video retrieval paths
  • Parallel search_batch MCP tool for multi-query workflows
  • explain_results MCP tool for search transparency and score inspection
  • Ingest-time image/video captioning to improve BM25 and keyword recall
  • Optional vision-language query expansion and stronger cross-modal benchmark coverage
  • PDF vision fallback so scanned or image-heavy documents remain searchable

Runtime and Product

  • HTTP transport for the MCP server with persistent model loading
  • Configurable model selection via environment variables and MCP config
  • Qwen3.5-0.8B captioning for lighter-weight media understanding
  • Version-aware server reporting and improved health/config surfaces
  • Collection management API for better multi-namespace organization

Reliability and Performance

  • Fast media retrieval defaults for better real-world latency
  • Query-side media handling fixes so reranking is no longer blind to media queries
  • Thread-safe model swaps plus explicit memory cleanup
  • Video frame sampling and cleanup fixes to reduce OOMs and tensor retention
  • Expanded UAT, observability, score audits, and release validation coverage
  • Trusted publishing workflow for tag-based PyPI releases

[0.1.0] — 2026-03-13

First public release. The local multimodal memory engine for AI agents.

Note: Python 3.12 and 3.13 supported. Python 3.14 is not yet supported due to pyarrow wheel availability.

Core

  • Cross-modal search across text, images, documents (PDF/DOCX/PPTX), and video in one unified query
  • Shared embedding space via Qwen3-VL (2048-dim vectors for all modalities)
  • 3-stage retrieval pipeline: embedding → reranking → query expansion (all multimodal)
  • Tiered search modes: embed (~1.7GB), hybrid (~3.4GB), full (~4.4GB) on MLX 4-bit

Backends

  • MLX backend for Apple Silicon with 4-bit and bf16 quantization
  • PyTorch backend for CUDA, MPS, and CPU
  • Auto-detection of best available backend
  • First-run model download UX with progress reporting

MCP Server

  • 17 MCP tools: search, search_fts, search_vec, ingest, index_document, index_image, index_folder, memory_add, memory_update, memory_delete, status, rebuild_fts, list_collections, list_namespaces, batch, get_config, set_config
  • Structured error responses with 4 standard codes
  • Batch operation support (up to 20 ops per call)
  • Runtime config inspection and adjustment

CLI

  • recallforge index — index files and folders
  • recallforge search — text, image, and video search
  • recallforge serve — MCP server mode
  • recallforge status — health check
  • recallforge watch — folder monitoring with start/stop/list/status

Storage

  • LanceDB + Tantivy FTS hybrid backend
  • Schema migration for namespace columns on existing stores
  • Content-addressable deduplication
  • Collection and namespace support

Performance

  • Warm search p50: 53ms (MLX 4-bit, Mac mini M4)
  • Warm search p95: 55ms
  • Cold start: 7.6s
  • Text indexing: 5.0 docs/sec
  • FTS miss fallback optimization (no expensive BM25 fallback on empty results)
  • Bulk mode context manager for deferred FTS rebuilds during batch ingest

Quality

  • COCO retrieval benchmark (50 images, MLX 4-bit embed mode):
    • Text → Image: R@1 23.6%, R@5 36.8%, R@10 46.4%
    • Image → Text: R@1 30.0%, R@5 42.0%, R@10 54.0%

Developer

  • 256+ unit tests, 35 SQL validation tests, comprehensive UAT suite
  • GitHub Actions CI (Python 3.12/3.13 matrix)
  • LRU embedding cache for repeat queries
  • Watch folder with path-keyed dedup and flush-on-shutdown
  • Configurable max file size on ingest
  • MIT License