Graph RAG

[dcieslak19973]

Is your feature request related to a problem? Please describe.
This is a half-baked idea right now, but the core idea is that while GPT-Researcher is in the research phase, that it would create/extract entity and relationship descriptions from the source documents and store those in a knowledge graph that would later be used in the content generation phase. See for example https://neo4j.com/blog/developer/graphrag-field-guide-rag-patterns/

Describe the solution you'd like

Adding support for Neo4j which combines graph databases with vector stores.

Describe alternatives you've considered
N/A

Additional context

See for example:

• https://neo4j.com/blog/developer/graphrag-field-guide-rag-patterns/
• https://github.com/microsoft/graphrag
• https://python.langchain.com/docs/how_to/graph_constructing/#llm-graph-transformer

[ElishaKay]

Sup @dcieslak19973

I like it!

The underlying abstraction we use for storing embedded content is the Langchain Vector Store.

There is a Neo4j Vector Store option we can leverage for this.

We'll want to follow the pattern of other swappable components of GPTR (Retrievers & LLM's) where the desired Vector Store provider can be set in the environment variables.

Green light for the PR

[ga-it]

200% on this!!

I'm struggling with GPTR and citations right now - I think it battles without a graph-RAG approach. Graph RAG would allow linking of source to knowledge.

We are running SciPhi R2R - https://r2r-docs.sciphi.ai/

It is very well suited to deep research:

It has python libraries - I have done some basic backend swapover for another app.