CodeWeave AI
Semantic codebase search and AI-powered code understanding — fully offline, completely free.
Ask anything about your codebase. CodeWeave indexes your project locally using Ollama and answers your questions by finding the most relevant code — no internet, no API key, no subscription.
✨ Features
- 100% local & offline — Your code never leaves your machine. No API keys, no cloud sync, no rate limits
- Natural language Q&A — Ask questions like “how does authentication work?” or “where is the retry logic?” and get accurate, source-linked answers
- Semantic search — Find code by meaning, not just keyword matching
- GraphRAG-powered retrieval — Understands function calls, imports, dependencies, and class hierarchies for deeper code awareness
- AI-powered explanations — Explain functions, trace execution flows, and summarize modules instantly
- @mention context pinning — Use
@filename.ts or @FunctionName to force-include specific files or symbols in context
- Clickable source citations — Every response links directly to the exact file and line range
- Streaming responses — Answers appear token-by-token in real time
- Multi-turn conversations — Remembers recent context for natural follow-up questions
- Auto re-indexing on save — Updated files are automatically re-embedded in the background
- Stale index detection — Detects when indexed files have changed and need refreshing
- Folder structure queries — Ask for project structure and get a real directory tree view
- Click-to-navigate — Jump directly from answers to source code locations
🚀 Getting Started
1. Install Ollama
Download and install Ollama from ollama.com/download, then start it:
ollama serve
2. Pull the required models
ollama pull qwen2.5-coder:7b
ollama pull mxbai-embed-large
First-time download is ~4–5 GB total. Only needed once.
3. Open a project and index it
Open any project folder in VS Code, then either:
- Click Index Codebase in the CodeWeave sidebar, or
- Run
CodeWeave: Index Codebase from the Command Palette (Cmd/Ctrl+Shift+P)
4. Ask questions
Type anything in the chat panel:
How does authentication work?
Where is the database connection initialized?
Explain the risk manager module
@retriever.ts what does the retrieve function do?
Show me the folder structure
💬 Example Questions
| Question |
What CodeWeave does |
How does auth work? |
Finds auth-related functions, traces the call graph |
Where are API calls made? |
Locates all HTTP fetch/axios calls across the codebase |
Explain the indexer |
Summarises the indexer module with code references |
@lancedb.ts what does search() do? |
Pins lancedb.ts and explains the search function specifically |
Show me the project structure |
Renders a real file tree from the workspace |
What calls the embed function? |
Uses the knowledge graph to trace callers |
📁 Custom Ignore Rules
Create a .codeweaveIgnore file in your project root (same syntax as .gitignore) to exclude specific files or folders:
# .codeweaveIgnore
secrets/
*.generated.ts
legacy/
migrations/
🔒 Privacy
CodeWeave is designed for privacy-first development:
- No telemetry — nothing is sent anywhere
- No cloud — all AI runs locally via Ollama
- No API keys — completely free, forever
- Local storage — the index is saved to
.codeweave-index/ inside your project
Add .codeweave-index to your .gitignore to avoid committing the index:
echo ".codeweave-index/" >> .gitignore
🏗️ How It Works
Your Code
│
▼
Walker ──► Entity Extractor ──► Ollama Embeddings
│
▼
LanceDB (local vector store)
+ Knowledge Graph (graph.json)
│
┌───────────┴────────────┐
│ │
Your Question Graph Traversal
(embedded) (calls/imports)
│ │
└───────────┬────────────┘
│
Retrieved Chunks
│
▼
Ollama Chat Model
│
▼
Streamed Answer
- Indexing — files are walked, parsed into entities (functions, classes, imports), embedded with
mxbai-embed-large, and stored in a local LanceDB vector database alongside a JSON knowledge graph
- Retrieval — your question is embedded, then the closest code chunks are found by cosine similarity; broad questions use graph-first traversal;
@mentions pin specific files or symbols
- Generation — retrieved chunks are assembled into a prompt and streamed through
qwen2.5-coder:7b
🤝 Contributing
Issues and pull requests are welcome on GitHub.
📄 License
MIT © parutechie
