RagLeap - No CUDA, no 2GB torch. 23MB reranker. 6 vector DBs.

The open-source core of RagLeap — a self-hosted RAG chat engine (document ingestion, retrieval, citations) powering RagLeap's full AI business manager: Voice AI, WhatsApp/Telegram/Discord bots, database & CRM integrations, workflow automation, and a private executive assistant. Foundation layer — hosted platform at ragleap.com - antonyrag/ragleap-core

I built this because every RAG library I tried assumed I had a GPU. RagLeap's reranker runs on ONNX Runtime — a 23MB quantized model instead of the 2GB+ torch/CUDA stack most RAG frameworks pull in even for CPU-only use. It supports 6 vector backends (pgvector, FAISS, Pinecone, Weaviate, Qdrant, Milvus), 8 embedding providers, and 12+ generation providers with automatic fallback if one goes down. Real per-call cost tracking with monthly budgets. 238 tests, real CI on every PR. Everything in the docs is either live-verified against real infrastructure or explicitly labeled unverified — no marketing claims, no numbers we haven't actually measured. pip install ragleap-rag Contributions welcome — the roadmap, and what we've deliberately said no to, are both documented in the open: github.com/antonyrag/ragleap-core Would love feedback, especially from anyone who's fought GPU dependency bloat before.

Quickstart for anyone who wants to try it:

pip install ragleap-rag

from ragleap import RagLeap, ProviderConfig, EmbeddingConfig

rag = RagLeap(

database_url="postgresql://user:pass@localhost/mydb",

embedder=EmbeddingConfig(provider="gemini", api_key="YOUR_KEY"),

primary=ProviderConfig(provider="gemini", api_key="YOUR_KEY"),

)

rag.init_schema()

rag.ingest("doc.pdf", open("doc.pdf","rb").read())

answer = rag.ask("your question")

print(answer["answer"])

Works on CPU with 23MB ONNX. No GPU needed.

I also added WhatsApp/Telegram/Discord examples in /examples folder on GitHub.

Which one should I demo next?

RagLeap - No CUDA, no 2GB torch. 23MB reranker. 6 vector DBs.

Replies