Advanced RAG: Exploring Query Rewriting

Sharing a article summary, I have read about RAG. Wonderful article by Florian. Some of these principles seem very natural and are part of normal conversation. Perhaps expanding those into various methodologies for LLMs.

Direct Link (Members Only) - https://medium.com/@florian_algo/advanced-rag-06-exploring-query-rewriting-23997297f2d1

Advanced RAG: Exploring Query Rewriting

This document focuses on query rewriting techniques within Retrieval-Augmented Generation (RAG) systems. RAG aims to improve the accuracy of responses generated by large language models (LLMs) by incorporating retrieval of relevant documents before generation. However, a key challenge is ensuring the semantics of the user's query align with the retrieved documents. Query rewriting techniques address this challenge by reformulating the original query.

Here's a breakdown of the key areas and findings explored in the document:

1. HyDE (Hypothetical Document Embeddings)

• Aims to bridge the semantic gap between query and documents by generating hypothetical documents that represent the user's intent.
• The LLM is used to create these hypothetical documents, which are then embedded alongside the original query for retrieval.
• While HyDE can improve retrieval quality, it may not always generate accurate hypotheticals if the LLM lacks knowledge on the topic.

2. Rewrite-Retrieve-Read

• This approach emphasizes rewriting the original query before retrieval.
• The LLM is used to reformulate the query to better suit information retrieval.
• This method is effective in improving the retrieved documents' relevance to the user's intent.

3. STEP-BACK PROMPTING

• This technique aims to improve the LLM's reasoning abilities for complex queries.
• It involves reformulating the query into a more abstract or broader question that is easier for the LLM to answer.
• The LLM then uses this broader understanding to answer the original, more specific query.
• This method is helpful for queries with intricate details that might hinder the LLM's reasoning process.

4. Query2Doc

• This approach focuses on query expansion using the LLM to generate pseudo-documents relevant to the query.
• These pseudo-documents are then combined with the original query to create a new, more informative query for retrieval.
• While not currently implemented in Langchain or LlamaIndex, it offers an alternative to HyDE's method.

5. ITER-RETGEN

• This method iteratively refines retrieval and generation within a Retrieve-Read-Retrieve-Read loop.
• It leverages the generated content from previous iterations to guide retrieval in subsequent iterations, aiming to improve the relevance of retrieved documents and the quality of the final response.

Overall, the document explores various query rewriting techniques that enhance the effectiveness of RAG systems. By reformulating user queries to better align with the document space, these techniques can significantly improve the accuracy and relevance of the information retrieved and ultimately the quality of the responses generated by LLMs.

?