/agentic

A minimalist Go implementation of Microsoft's GraphRAG

Primary LanguageGoMIT LicenseMIT

agentic

This package serves as a workspace for exploring state of the art RAG approaches in Go. The initial focus is on the GraphRAG approach from Microsoft, but other approaches may be explored in the future.

Areas of Exploration

  • GraphRAG
  • dsRAG

Findings

GraphRAG

GraphRAG is seems to be mostly marketing spin from Microsoft. There are some legitimately cool ideas in there such as using LLMs to extract graph relationships from a given text.

The heavy use of multi-stage summarisation, however, produces results that are not accurate to the original text or useful for an executive or research audience because they too general and lacking concrete details.

dsRAG

Source: https://github.com/D-Star-AI/dsRAG

There are three key methods used to improve performance over vanilla RAG systems:

Semantic sectioning AutoContext Relevant Segment Extraction (RSE)

We will implement these in Go to explore them further.