Caliby: Embedded Vector Database for LLM Agents
Integrate Caliby via pip install caliby and replace your current vector DB for agent apps.
Integrate Caliby into your agent stack by running pip install caliby and migrating your vector storage to the new library.
Summary
Caliby is an open‑source, pure‑embedded vector retrieval library co‑developed by Sea‑Land AI and MIT’s Michael Stonebraker team. It is written in C++ with Python bindings and can be installed with a single pip command. Caliby supports HNSW, DiskANN, and IVF+PQ indexes, and it stores text and vectors together in a unified buffer pool. The library is CPU‑SIMD accelerated (AVX‑512/AVX2/SSE) and offers multi‑threaded parallel retrieval and prefetch optimisations. It can persist indexes to disk without any external services, making it ideal for agent and RAG workloads that run in‑process. Benchmarks show Caliby builds 4–5× faster than pgvector and delivers 5× higher query throughput at 90 % recall. The library is released under an open‑source licence and is available on GitHub at https://github.com/zxjcarrot/caliby. Developers can drop Caliby into their existing Python code and replace heavier vector backends with a single in‑process dependency.
Key changes
- Caliby is a pure‑embedded vector DB with HNSW, DiskANN, and IVF+PQ indexes
- It stores text and vectors together in a unified buffer pool
- CPU‑SIMD acceleration (AVX‑512/AVX2/SSE) and multi‑threaded parallel retrieval
- Disk persistence without external services
- Build speed 4–5× faster than pgvector
- Query throughput 5× higher at 90 % recall
- Single pip install with no dependencies
- Available on GitHub at https://github.com/zxjcarrot/caliby