Faceted RAG - semantic clustering

[wassfila]

This discussion focus on combining the text search faceting feature with a Retrieval Augmented Generation system

Analysis

RAG usually assume the capability of flawless retrieval similarities matching and assume the idea that the user is interacting with a know-it-all system. This discussion steps back and downgrades expectations from RAG system, and therefore brings back the user in the loop of sorting information sources and scope. The granularity goes down to faceting the content and exposing it to the user during a first search phase before invoking the LLM to summarize the search results.

Survey of 'Faceted RAG' similar strategies

• Augmented and Adaptive Knowledge Retrieval, Izacard and Grave, 2021 https://arxiv.org/pdf/2007.01282
• Fine-grained 'Topic Modeling' for RAG
• Interactive RAG Systems

Manual and Automated faceting

• The manual faceting can be part of the documents construction phase where they naturally come from different sources, so the source meta data can help characterize the documents
• Automated faceting can leverage clustering techniques combined with summary generation and auto cluster-label definition techniques

[wassfila]

this is more fairly described as

Semantic clusturing