RAG chunking can make retrieval worse and force costly rework
A learning RAG app was built with split by and stored in Qdrant. During splitting, code markers such as '@tool' disappeared, and the resulting chunks did not match the way needed to work. Loading the full set into Qdrant took about 48 minutes because local embedders were used.
When started, the best chunks were not ranked well, so the app could not reliably find the right context. The central issue is how to choose a better chunking method and how to replace chunks that are already loaded in Qdrant.
Key points
- was split with before being stored in Qdrant.
- Important code markers such as '@tool' disappeared during chunking.
- Full took about 48 minutes with local embedders.
- Poor chunking led to weak ranking.
- Testing a small sample first can prevent expensive rework.