Going Meta - Ep 22: RAG with Knowledge Graphs

Episode 22 of Going Meta - a series on graphs, semantics and knowledge Jesús Barrasa: / barrasadv
Links from the Show:
Vector Search: bit.ly/4cuZoeS
Educational Chatbot: bit.ly/45RvnDz
Structure Aware Retrieval: / adding-structure-aware...
GenAI Walkthrough:
bit.ly/4cxfPXW
GenAI App Building: bit.ly/4cxfPXW
DevOps Rag Application: / using-a-knowledge-grap...
LangChain: github.com/langchain-ai/langc...
0:00 Welcome
6:35 Recap on Data Semantics
11:28 RAG
20:40 Knowledge Graphs to improve RAG
31:11 Q&A
36:25 Code Example
55:50 More Q&A
1:00:55 WrapUp
Repository: github.com/jbarrasa/goingmeta
Knowledge Graph Book: bit.ly/3LaqE6b
Check out community.neo4j.com/ for questions and discussions around Neo4j
#neo4j #graphdatabase #knowledgegraphs #knowledgegraph #semantic #ontology #rag

[@alimahmoudmansour9681]

Great... thenks a lot

[@neo4j]

You're very welcome!

[@AdamLorentzen]

This was so helpful, thank you so much!!! I still don't understand how the LLM knows what the Nodes are and how they are related, especially for a company with their own taxonomy. Do you have to pass that info to the LLM to provide context? Or does the Lanchain RAG functions inherently do that? Thanks, great series!

[@jbarrasa4649]

In our case, the KG offers you a `pathsim.search` method that is taxonomy-aware. So if you store your taxonomy in your KG in a standard way, then you can leverage it for "graph semantic similarity" using the available functions (like `pathsim.search` and others) or even through custom exploration. That's the retrieval part of the RAG pattern, and therefore the LLM does not need to be aware of it. All the LLM receives is the result of the exploration in the graph in the form of context. I hope it makes sense?

[@dattashish]

Informative ! though it would be nice if the screen resolution was as good as your photos 🙂 The graph and LLM seem to be too intertwined to get it to work. Maybe you should try to create a toolkit to ease thing for the users for the entire pipeline required.

[@neo4j]

Sorry - we should have zoomed in a bit more!

[@Tortilla_Jankins]

which you can totally run on Neo4J btw :)

[@vivalancsweert9913]

This was very interesting and inspiring! Thank you!! Where is the discord channel?

[@neo4j]

glad you liked it! You can join us on discord: discord.gg/neo4j

[@sahil0094]

Can we create one vector index for multiple node labels? If not then should we create vector index of each label and how do we query on multiple vector index then?

[@neo4j]

Not yet. A vector index currently covers a single label. The simplest workaround is to create a new label dedicated to scoping your vector index. Querying multiple indexes and combining the results is possible using a pattern like: CALL { CALL db.index.vector.queryNodes('vIndex1', 5, $query) YIELD node, score RETURN node, score UNION CALL db.index.vector.queryNodes('vIndex2', 5, $query) YIELD node, score RETURN node, score } RETURN node, score ORDER BY score DESC