Fine-tuned RAG reports an 11% gain in retrieval hit rate
A lightweight changes a before a RAG system searches for supporting passages. Standard gives every embedding dimension equal weight, so wording that looks similar can outweigh information that is truly relevant. The new approach learns which dimensions matter for a particular document collection and reduces the effect of noisy ones.
To create , a writes several answerable questions for each document chunk. The question-chunk pairs are converted into and split into training and validation sets. After each training round, validation questions test how often the system retrieves the correct chunk.
The reported results are an 11% increase in retrieval hit rate, a 12% increase in answer completeness, and a 9% increase in to the supporting material.
Key points
- A transforms each before retrieval.
- The comes from multiple LLM-generated questions for every document chunk.
- Question-chunk pairs are divided into separate training and validation sets.
- The correct-chunk retrieval rate is checked after every training round.
- Reported gains are 11% in hit rate, 12% in completeness, and 9% in .