Quick Solution to fix RAG

1. You need to trace immediately

If you are not tracing your project, I would suggest to tracing immediatedly using LangSmith, and LangSmith comes with the built in evaluators to evaluate your document retrieval and even more using LLM as Judge option

2. Now that you know what is makes good retrieval

Understand that every data is unique, we cannot say on first look. I would strongly encourage you to create a dataset in LangSmith and test with multiple version. By this time you would have a rough idea of which works