Improve RAG performance on custom vocabulary | by Teemu Sormunen | Feb, 2024

Summary

The article discusses the challenges of using semantic search models, especially when dealing with custom vocabulary, as seen in the context of error codes like "APP_w304". The author critiques traditional semantic search techniques and suggests supplementary methods like BM25 (keyword search), query augmentation, and Reciprocal Rank Fusion (RRF) to improve search performance. Fine-tuning of large language models with appropriate training data generation is also presented as a solution to align keyword and semantic search results for more efficient information retrieval within documents, such as a Tesla Model 3 manual.