If you’ve ever spent hours tweaking a Retrieval-Augmented Generation (RAG) pipeline—wondering whether BM25 or a vector index works better, or if duplicating a passage in your prompt helps—AutoRAGLearnings is here to save you time. This toolkit:

  1. Turns your docs into Q&A
  2. Chunks & embeds content (locally via PGVector or in your Azure Search index)
  3. Greedily tests each RAG step to lock in the best module by measuring context precision
  4. Lets you ask any question with one simple command

Grab the full code on GitHub:

https://github.com/jig21nesh/myautorag

Why Greedy Search?

Manually testing every combination of RAG modules is both tedious and time-consuming:

  • Does BM25 outperform a dense vector store?
  • Would reranking with an LLM beat a simple pass-through?
  • Should I tweak my prompt builder or stick with f-strings?

Greedy search cuts straight to the chase. Instead of exploring all pipelines, it:

  1. Iterates node-by-node (query_expansion → retrieval → augmentation → reranker → prompt_maker → generator)
  2. Swaps in each candidate in isolation
  3. Measures context_precision on a ground-truth Q&A set
  4. Lock in the winner before moving on

That reduces tests from “all combinations” to the sum of candidates per node—orders of magnitude fewer runs, yet still near-optimal.

I hope AutoRAGLearnings helps you tune your RAG workflows in minutes instead of days. Give it a try, star the repo, and leave a comment