Translations:FACTS About Building Retrieval Augmented Generation-based Chatbots/68/en

ChipNemo (10) presents evidence for using a domain adapted language model for improving RAG’s performance on domain specific questions. They finetuned the e5-small-unsupervised model with 3,000 domain specific auto-generated samples. We tried fine-tuning e5-large embeddings model in Scout Bot. Our results did not demonstrate significant improvements. We are presently collecting high quality human-annotated data to repeat the experiments. This could be an important direction to explore in the future for our work. Another interesting technique was presented by Setty et. al. (15), in improving RAG performance using Hypothetical Document Embeddings (HYDE) technique. HyDE uses an LLM to generate a theoretical document when responding to a query and then does the similarity search with both the original question and hypothetical answer. This is a promising approach but might make the architecture complex.