Here are key strategies to improve RAG systems when dealing with large document collections:
Proper document chunking is crucial for effective retrieval:
-
Use semantic chunking instead of fixed-size chunking to preserve context
We use fixed size chunking - but it wont cut trough midle of a sentence. It uses NLP to identify sentence boundaries
-
Experiment with overlap between chunks (typically 10-20%)