Knowledge Share
Technical articles, tutorials, and insights
Open Source Project of the Day (#159): Vibe-Trading — Quantitative Research in Natural Language, an AI-Powered Personal Trading Agent
HKUDS's open-source AI trading research framework turns natural language questions into runnable quantitative strategies, backtests, and trading workflows. 68 agent tools, 19 free data sources, 461 pre-built alpha factors, multi-market backtest engines, shadow account behavioral diagnostics. 21.8k Stars, MIT license.
Open Source Project of the Day (#156): OpenDataLoader PDF — The #1 Ranked Open-Source PDF Parser, Built for RAG
OpenDataLoader PDF by Hancom and veraPDF developers scores 0.907 overall and 0.928 on table extraction across 200 real-world PDFs — top of open-source benchmarks. XY-Cut++ reading order algorithm, built-in AI safety filtering, automatic accessibility tagging. LangChain native integration, runs locally with no GPU. 26.8k Stars, Apache-2.0.
Open Source Project #136: OpenMemory — A Real Cognitive Memory Engine for AI Agents
Self-hosted cognitive memory engine for AI agents with five memory sectors (episodic, semantic, procedural, emotional, reflective), temporal knowledge graph, and Waypoint association graph. 110-130ms query latency, 6-10× cheaper than SaaS alternatives. Python/Node SDK, MCP, LangChain/CrewAI support. Apache 2.0.
Agent Series (10): MCP Protocol — Standardizing the Tool Ecosystem
A deep dive into how Model Context Protocol (MCP) solves the tool ecosystem management problem. Three runnable demos cover traditional Function Calling limitations, MCP Server dynamic tool discovery, and LLM agents using MCP tools via langchain-mcp-adapters. Includes Host/Client/Server architecture breakdown and a complete MCP vs Function Calling comparison.
RAG Series (18): Conversational RAG — The Pronoun Problem in Multi-Turn Dialogue
Part 18 of the RAG series. Single-turn RAG sends each raw question straight to retrieval. But follow-up questions like "what are its four metrics?" and "which one is hardest to improve?" fail completely without context. History-Aware Retriever rewrites each follow-up into a standalone question before retrieving. The experiment also reveals an interesting reversal — RAGAS metrics showed no advantage for ConvRAG, but the qualitative retrieval comparison tells the real story.
RAG Series (13): Query Optimization — Asking Better Questions
Part 13 of the RAG series. Why does vector retrieval swing wildly based on phrasing? How does Multi-Query use multiple angles to widen recall? Why does HyDE search with a "fake answer" instead of the question? How does Query Decomposition break down complex questions? RAGAS results across four strategies: context_recall improves from 0.625 to 0.875. Full code included.
RAG Series (11): Rerank — Putting the Right Documents First
Part 11 of the RAG series. How bad is vector search ranking quality? What's the real difference between Bi-Encoder and Cross-Encoder? How do you wire up a Reranker with ContextualCompressionRetriever? We measure with RAGAS: context_precision jumps from 0.552 to 0.792. Full LangChain implementation included.
RAG Series (12): Advanced Chunking — Parent-Child and Contextual Retrieval
Part 12 of the RAG series. What's the fundamental flaw in naive chunking? How does Parent-Child use small chunks for search and large chunks for context? What does Anthropic's Contextual Retrieval add to each chunk? RAGAS results: context_recall improves from 0.625 to 0.875, context_precision from 0.583 to 0.938. Full LangChain implementation included.
RAG Series (10): Hybrid Search — Retrieving More, Missing Less
Part 10 of the RAG series. Why does pure vector search fail on exact keyword queries? What are BM25 and vector search each good at? How does the RRF fusion algorithm work? We run 6 test queries and use MRR to quantify the difference between three retrieval strategies. Full LangChain EnsembleRetriever code included.
RAG Series (4): Document Processing — From Raw Files to High-Quality Chunks
Part 4 of the RAG series. A deep dive into 4 chunking strategies — Fixed Size, Recursive Character, Semantic Chunking, and Document Structure. Side-by-side comparison using the same Markdown document, with visualizations and a selection decision matrix.
RAG Series (2): Building Your First RAG Pipeline with LangChain
The second article in the RAG series. Building a production RAG pipeline with LangChain 1.x, ChromaDB, and SiliconFlow. Full source code available on GitHub.
Open Source Project of the Day (Part 23): PageLM - Open-Source AI Education Platform, Turning Learning Materials into Interactive Resources
A deep dive into PageLM, an open-source AI education platform by CaviraOSS inspired by NotebookLM — converting PDFs and documents into quizzes, flashcards, notes, and podcasts, supporting multiple LLMs and TTS engines, ideal for students, teachers, and researchers