I Increased Retrieval From Top-5 to Top-20. My Answers Got Worse

1. 1Retrieve top-20 candidates using a vector database like Faiss or Pinecone
2. 2Filter the retrieved candidates using a ranking model or a simple heuristic like BM25
3. 3Evaluate the impact of increased retrieval on answer quality using metrics like accuracy or F1-score
4. 4Adjust the filtering threshold or ranking model to optimize answer quality
5. 5Test the updated RAG system with a new set of questions to validate the improvements

This lesson is beneficial for NLP engineers and data scientists working on question-answering systems, as it highlights the importance of filtering in RAG retrieval

💡 More retrieval candidates do not always lead to better answers, proper filtering is crucial

🤖 Increasing RAG retrieval candidates from top-5 to top-20 can lead to worse answers if not properly filtered 📊