# Multi-Meta-RAG: Improving RAG for Multi-Hop Queries using Database Filtering with LLM-Extracted Metadata

**Authors:** Mykhailo Poliakov, Nadiya Shvai  
**arXiv:** [2406.13213](https://arxiv.org/abs/2406.13213)  
**Date:** 2024-06-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.13213v2/

## Key points

- Multi-Meta-RAG method enhances retrieval-augmented generation (RAG) for multi-hop queries using database filtering and large language model (LLM)-extracted metadata
- Results show significant improvements in chunk retrieval and LLM generation compared to Graph RAG
- Limitations include the need for specific domain and question format queries, manual creation of prompt templates, and results falling short of feeding LLM precise ground-truth facts
- Evaluation results compare GPT-4 and Google PaLM performance on different query types, with both models achieving high accuracy scores above 0.9 for inference queries
- Future plans involve exploring more generic prompt templates for metadata extraction, testing alternative LLMs like LLama 3.1 on datasets with recent cut-off dates

## Summary

, , , , 

The authors of this study introduce Multi-Meta-RAG, a method that utilizes database filtering with large language model (LLM)-extracted metadata to enhance retrieval-augmented generation (RAG) for multi-hop queries. The results of experiments demonstrate that Multi-Meta-RAG significantly improves chunk retrieval and LLM generation compared to alternative solutions such as Graph RAG. However, there are limitations to this approach, including the need for specific domain and question format queries for metadata extraction, manual creation of prompt templates, and results that still fall short of feeding LLM precise ground-truth facts. The evaluation results in Table 4 compare the performance of GPT-4 and Google PaLM on different types of questions. Both models achieve high accuracy scores above 0.9 for inference queries, with Google PaLM outperforming GPT-4 in comparison and temporal queries but struggling with Null questions. The authors suggest that combining both models for different query types could further improve overall accuracy. In the future, more generic prompt templates will be explored for metadata extraction using multi-hop datasets from various domains, and alternative LLMs like LLama 3.1 will be tested on datasets with more recent cut-off dates. Despite its limitations, Multi-Meta-RAG shows promise in enhancing RAG performance for multi-hop queries and offers a relatively straightforward and explainable solution compared to other methods. This research was partially funded by the OpenAI Researcher Access Program (Application 0000005294). Additionally, an Appendix provides a Metadata Extraction Prompt Template for extracting metadata from questions to filter database sources related to article content. The template includes XML feed data related to the study's topic on Multi-Meta-RAG.