This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+14 more
Summary
AI summaryModels cache persists across reboots, new exclude patterns filter unwanted files during indexing.
Full changelog
What's New
Persistent Model Cache (models_cache_dir)
FastEmbed ONNX models (~250MB) now persist across reboots. No more re-downloading after every restart on Linux.
Exclude Patterns (exclude_patterns)
Skip directories and files during indexing with glob patterns. Perfect for node_modules/, .venv/, .git/, and other junk.
Jupyter Notebook Parser (.ipynb)
Dedicated parser extracts only markdown and code cell sources — no more indexing base64 images, execution counts, or cell metadata.
MCP stdio Protection
stdout redirected to stderr before mcp.run(), preventing stray print() calls from corrupting the JSON-RPC stream.
inotify Resilience (Linux)
File watcher gracefully degrades when Linux inotify limits are exceeded — server continues without auto-reindex instead of crashing.
MetaTrader Support (.mqh / .mq4)
MQL4/MQL5 files now parseable as code. Not in defaults — enable via supported_formats in config.
Configuration
paths:
models_cache_dir: "./models_cache" # persistent ONNX cache
documents:
exclude_patterns:
- "node_modules"
- ".venv"
- "__pycache__"
- ".git"
Tests
47 tests passing (11 new), zero failures. Covers all new features with edge cases.
Install / Upgrade
pip install --upgrade knowledge-rag
Community Credit: Ideas for models_cache_dir, exclude_patterns, .ipynb support, and inotify handling originated from @Hohlas (PR #18). Implementation rewritten from scratch with proper architecture and rigorous tests.
Full Changelog: https://github.com/lyonzin/knowledge-rag/compare/v3.3.2...v3.4.0
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About lyonzin/knowledge-rag
Local RAG system for Claude Code with hybrid search (BM25 + semantic), cross-encoder reranking, markdown-aware chunking, query expansion, and 12 MCP tools. Runs entirely offline with zero external servers.
Related context
Beta — feedback welcome: [email protected]