Improper chunking technique degrades retrieval performance

When chunking techniques result in chunks with too little context, oddly split chunks, or too large chunks, retrieval performance suffers. This is one of several root causes that should be investigated before resorting to fine-tuning an embedding model.

Examine a few chunks manually to evaluate whether the current chunking technique is appropriate. Experiment with alternative chunking techniques such as late chunking. Consider keyword search or hybrid search if many queries require exact keyword matches.