Witchcraft: Fast Multi-Vector Semantic Search Engine Released in Rust with Pickbrain Integration
Key Takeaways
- ▸Witchcraft delivers 21ms p.95 search latency and 33% NDCG@10 accuracy—twice as fast as the original XTR-Warp implementation—while running entirely on-device from a single SQLite database
- ▸The engine requires no external vector databases, API keys, chunking strategies, or complex infrastructure, making it suitable for immediate deployment in production applications
- ▸Pickbrain demonstrates practical applications by indexing AI coding session transcripts, enabling developers to maintain persistent memory across multiple Claude Code and Codex sessions with fast semantic recall
Summary
Witchcraft, a from-scratch Rust reimplementation of Stanford's XTR-Warp semantic search engine, has been released as an open-source project offering exceptional performance and ease of deployment. The engine achieves 21ms p.95 end-to-end search latency on NFCorpus with 33% NDCG@10 accuracy—more than twice as fast as the original XTR-Warp on server-class hardware—while running entirely from a single SQLite database with no external dependencies, vector database, or API keys required.
The project includes Pickbrain, a practical sample application and agent skill that enables semantic search across Claude Code and OpenAI Codex session transcripts, effectively providing AI agents with persistent long-term memory and the ability to search across all historical coding sessions. The combination allows developers to instantly query questions like "what was that conversation where I fixed the auth middleware?" across their entire project history. Witchcraft is designed for client-side deployment with multiple backend options optimized for different hardware platforms, including support for Apple Silicon GPU acceleration, Intel x86 optimization, and OpenVINO inference.
- Multiple hardware-optimized backends support Apple Silicon, Intel x86, Windows, and Node.js environments, allowing flexible integration across different development platforms
Editorial Opinion
Witchcraft represents a significant advancement in making semantic search more accessible and practical for developers. By combining the performance advantages of a Rust implementation with the simplicity of SQLite-based storage, it challenges the assumption that semantic search requires complex infrastructure. The Pickbrain use case is particularly compelling—offering AI agents genuine long-term memory across coding sessions could fundamentally change how developers interact with AI coding assistants, transforming them from stateless conversation tools into project-aware collaborators with contextual continuity.
