How Do I Evaluate Chunking Strategies for Rags

Image from Lummi.ai

I’ve researched RAGs extensively and know that chunking is critical to any RAG pipeline.

Many people I’ve talked with trust that better models could improve RAGs. Some put too much trust in vector databases. Even those who agreed that chunking is important didn’t think it could significantly improve the system.

Most of them argue that large context windows would replace the need for chunking strategies.

But chunking techniques are here to stay. They are effective and a must for any RAG project.

However, a key question remains unanswered: How can I pick the best chunking strategy for a project?

In the past, I’ve discussed several strategies: recursive character splitting, semantic chunking, and agentic chunking, and even argued clustering as a fast and cheap alternative to agentic chunking.