ragleap-rag
CPU-only RAG library — no GPU, no 2GB torch stack
Every RAG library I tried had the same problem: reranking needed
torch + CUDA, which meant a 2GB+ dependency even if all you wanted
was CPU inference on a small server.
So I built ragleap-rag around a 23MB ONNX quantized reranker instead.
No GPU required, no massive install.
It grew from there into a fuller RAG toolkit:
- 6 vector backends (pgvector, FAISS, Pinecone, Weaviate, Qdrant, Milvus)
- 8 embedding providers, 12+ generation providers with automatic fallback
- Real per-call cost tracking with monthly budgets and auto-downgrade
- 238 tests, CI on every PR
It's fully open source: pip install ragleap-rag
I'm a solo dev building this alongside a commercial SaaS platform, so
it's genuinely a side project - built because I needed it, not to sell
anything. Would love feedback, especially from anyone who's hit the
same GPU-dependency wall.
1 Comment
Happy to answer anything - architecture decisions, why ONNX over
sentence-transformers, roadmap, whatever. Also open to hearing what
would make this more useful for your stack if you've built RAG
pipelines before.