Find which floor of your RAG is wrong before you swap the model
A RAG that answers badly has two floors that can break: retrieval and generation. Five questions, asked in order ([Instagram reel Dd6q1k9FGZU](https://www.instagram.com/reel/Dd6q1k9FGZU/)), tell you which one, and which fix to try first.
Why
When your RAG answers badly, the model is rarely the culprit. You swap the model, rewrite the prompt, add documents, and the bad answer comes back the next day in another shape. You touched the end of the chain while the fault sat at the start: the right passage was never retrieved, or it was cut off before reaching the model. This kit splits the two floors. Retrieval: did you find the right chunks, and all of them? Generation: did the model use them properly? Each question has its own measure and its own fix. A RAG is debugged like plumbing: you find the leak by following the pipe up, not…
When
Phase: you have a RAG running on your documents and at least a dozen questions whose right answer you know. Audience: a developer or PM who must say where it breaks. ✅ You know it's your kit when The answer is wrong but plausible: the model fills a retrieval gap You changed the prompt three times with no measurable gain You have to defend your RAG in front of someone and all you have is…
Tools included
- Llm Evals
- Reranking
- Chunking
- Recherche Hybride Rrf
- Embeddings
- Deepeval