Knowledge Share
Technical articles, tutorials, and insights
Open Source Project #172: agentOS — An Operating System as a Library for AI Agents, 92x Faster Cold Starts, 47x Less Memory
AI agent runtime embedded in your Node.js backend as a library. WebAssembly + V8 isolates, no microVMs/containers/sandboxes needed. Cold start 4.8ms (vs E2B's 440ms), 22MB per instance (vs Daytona's 1GB). Built-in Pi, Claude Code, Codex, OpenCode. Supports filesystem mounts, host bindings, cron, multiplayer, agent-to-agent delegation. Rust + TypeScript, Apache 2.0, 4.3k Stars.
Open Source Project #173: AI For Beginners — Microsoft's Complete 12-Week AI Curriculum, 55k Stars
Systematic AI beginner curriculum from Microsoft. 12 weeks, 24 lessons covering symbolic AI, neural networks, computer vision (CNN/GAN/object detection/semantic segmentation), NLP (word embeddings/RNN/Transformer/BERT/LLM), genetic algorithms, deep reinforcement learning, multi-agent systems, and AI ethics. Each lesson includes theory + Jupyter Notebook + lab. Supports PyTorch and TensorFlow dual paths. 55.7k Stars, MIT, 50+ language translations.
Codebase Knowledge Base (10): Incremental Updates — When to Rebuild the Index, and Which Parts
Code never stops changing, and you can't do a full index rebuild on every commit. Starting from a real detect_changes output, this article builds a three-tier decision tree: first determine whether the change affects retrievable content, then pinpoint exactly which functions changed, then use FILE_CHANGES_WITH edges to surface hidden temporal coupling. Rebuild only what truly needs rebuilding — compress incremental update cost to a fraction of full rebuild.
Codebase Knowledge Base (11): Cross-Repo Scenarios — When One Service Calls Another
A single-repo knowledge base can answer 'where is this function, who calls it, how should I change it.' But real engineering systems are multiple repositories collaborating over HTTP, message queues, and gRPC. Starting from a real cross-repo-intelligence run — LightRAG × graphrag returns 0 edges — this article explains why that result is correct, and what cross-repo analysis actually finds: integration relationships, not functional similarity.
Codebase Knowledge Base (12): Git History Is the Fourth Retrieval Path
The vector path answers 'what is this,' the graph path answers 'who calls it,' the symbol path answers 'where is it.' But engineers often need to ask a fourth type of question: 'why is this code written this way?' — and the answer to that question isn't in the code. It's in git history. Starting from LightRAG's 465 FILE_CHANGES_WITH edges, this article shows how to turn git history into a queryable fourth retrieval path.
Codebase Knowledge Base (13): Evaluation — How to Know Your Knowledge Base Is Good Enough
The knowledge base is built. Now what? Recall@5 is the go-to metric for RAG evaluation, but codebase retrieval has fundamentally different requirements from document RAG: finding a relevant paragraph is 'good enough' for a blog search, but code retrieval demands function-level precision. This article designs a purpose-built evaluation framework for codebase knowledge bases: four metrics, a dataset construction methodology, and how to wire evaluation into CI/CD to continuously track knowledge base quality.
Open Source Project of the Day (#171): Harness Handbook — Give AI Agents a Navigable Behavior Map Before Asking Them to Edit Agent Code
Companion tool to arXiv:2607.13285 (July 2026). Converts any AI agent harness codebase into a behavior-centric navigable handbook using static program analysis + LLM structuring. Three-level document tree + state-register view + BGPD algorithm guides agents from behavioral descriptions to precise code locations. Tested on Codex (Rust, 2,267 files): 45.6% win rate vs 26.7% baseline, 12.7% token reduction. 252 Stars, Apache-2.0.
Open Source Project of the Day (#169): AIRI — Self-Hosted AI VTuber That Actually Plays Minecraft
Open-source, self-hosted AI virtual companion system inspired by Neuro-sama. Web-first architecture built on Vue.js + WebGPU with realtime voice chat, Live2D/VRM avatars, Minecraft/Factorio game agents, and 30+ LLM provider support. Spawned the xsai and unspeech sub-projects. 44.8k Stars, MIT license.
Open Source Project of the Day (#170): CodeWiki — ACL 2026 Research-Grade Codebase Documentation with Recursive Multi-Agent Architecture
FPT Software's open-source codebase-level documentation generation framework, accepted at ACL 2026. Tree-Sitter AST parsing + recursive multi-agent architecture + hierarchical synthesis, handling codebases from 86K to 1.4M lines across 9 languages. Outputs Markdown docs with Mermaid architecture diagrams. Benchmarked above DeepWiki (68.79% vs 64.06%) on CodeWikiBench. 1.5k Stars, MIT license. Includes comparison with similar open-source projects.
Codebase Knowledge Base Series (05): Vector Retrieval vs Knowledge Graph — Adding a Call Graph Didn't Help?
Vector retrieval (Recall@5=0.958) vs graph-augmented retrieval (seed_k=3, BFS 2 hops): identical overall scores, but different failure modes. Graph expansion fixes Q8 (calculate_order_total reached via process_checkout in 2 hops) but breaks Q1 (expanded candidate pool crowds out verify_password). Conclusion: naive graph traversal is double-edged; the real engineering value is encoding structural information into embedding content, not bolting on graph traversal at retrieval time.
Codebase Knowledge Base Series (06): Encoding the Call Graph into Embeddings — Structural Augmentation Works, But Not Enough
Testing the hypothesis from Article 05: encoding called_by/calls into embedding text to fix Q8 (calculate_order_total missing). Result: structural prefix raises calculate_order_total's similarity score from 0.46 to 0.51 — right direction, but other payment module functions rank higher, blocking it from top-5. Strategy B (comment prefix) also causes Q7 regression. Strategy C (docstring injection) is stable but yields no improvement. Conclusion: the semantic gap is a fundamental embedding limit, not a workaround-able implementation detail.
Codebase Knowledge Base Series (07): Hybrid Search BM25 + Vector — Q8 Still Fails, and Total Score Drops
Hybrid search (BM25 + vector RRF fusion) is widely considered best practice for vector retrieval augmentation. It should fix Q8. Result: calculate_order_total has zero token overlap with the query — BM25 cannot find it at all. Hybrid inherits BM25's Q7 regression, dropping overall Recall@5 from 0.958 to 0.931 — worse than pure vector. Five articles have now systematically exhausted all text-matching variants. Conclusion: Q8's root cause is in code structure, not text; the pure text-matching route has reached its limit.