Over the past few months, while building RAG applications, I kept noticing the same pattern.
Every new project meant rebuilding the same pipeline from scratch:
- choosing an embedding model
- selecting a vector database
- experimenting with chunking strategies
- switching between LLMs
- adding rerankers
- integrating observability
- evaluating responses
- wiring it all together in code
The business logic changes across projects, but the underlying RAG architecture is largely the same. Yet every team ends up re-solving it, and optimization always gets pushed to "later."
So I built NexRAG — an open-source framework that lets you build production-ready RAG pipelines using simple YAML configuration instead of rewriting orchestration code every time. What used to take 3,000+ lines of custom code now takes roughly 300 lines to get a production-grade pipeline wired up.
With a few config changes, you can:
- Swap LLM providers
- Change embedding models
- Try different chunking and retrieval strategies
- Switch vector databases
- Configure rerankers
- Enable LLM-as-a-Judge evaluations
- Add observability and evaluation metrics
- Build custom components and plug them into the pipeline
The idea is simple: configuration over boilerplate.
Need something custom? Every component is extensible. Build your own retriever, embedder, chunker, vector store, or evaluator, and plug it in — no changes to the core framework required.
From day one, NexRAG is designed with production in mind — not just retrieval, but observability, evaluation, and security, which most AI projects only think about after something breaks.
🔗 NexRAG: https://lnkd.in/d4QRBZmb
🔗 Example project (Simple Policy Bot): https://lnkd.in/drgMSq4C
This is just the beginning. NexRAG is fully open source, and I'd love to collaborate with developers working in AI infrastructure, RAG, LLMOps, and production AI systems. Feedback, issues, feature requests, and contributions are all welcome.
Looking forward to building this together.
Top comments (0)