ContextWeaver addresses the challenge of providing AI agents with precise, contextually rich code snippets from local repositories. It targets AI code assistants and developers building LLM-powered coding tools, enabling deeper code understanding and more effective AI-driven development through advanced semantic retrieval.

How It Works

ContextWeaver employs a hybrid search engine combining vector (semantic) and lexical (FTS) retrieval, fused via Reciprocal Rank Fusion (RRF). Code is semantically split using Tree-sitter AST parsing for supported languages (TS, JS, Python, Go, Java, Rust), with a fallback for others. A three-stage context expansion strategy (neighbor, breadcrumb, import) enriches retrieved snippets before token-aware packing and truncation.

Quick Start & Requirements

• Installation: Global installation via npm/pnpm (npm install -g @hsingjui/contextweaver or pnpm add -g @hsingjui/contextweaver).
• Prerequisites: Node.js >= 20. Requires configuration of external Embedding and Reranker API keys and endpoints in ~/.contextweaver/.env after running contextweaver init.
• Usage: Indexing is initiated with contextweaver index [/path/to/repo]. Semantic search is performed via cw search --information-request "...". MCP server activation uses contextweaver mcp.

Highlighted Details

• Hybrid search (vector + FTS) with RRF fusion for robust retrieval.
• AST-based semantic code splitting via Tree-sitter for 6 primary languages.
• Multi-stage context expansion (neighbor, breadcrumb, import) for comprehensive context.
• Native MCP server support for seamless AI agent integration.
• Smart TopK truncation strategy optimizes context length while ensuring recall.

Maintenance & Community

The provided README does not detail specific contributors, community channels (e.g., Discord/Slack), or roadmap links.

Licensing & Compatibility

Distributed under the permissive MIT License, suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

Relies on external API keys for embedding and reranking, necessitating setup and potential costs. Unsupported languages use a less precise line-based fallback splitting mechanism.