Got interested in knowledge graphs and expanding my understanding of the RAG landscape more as I was thinking about the design of Honcho.
This is a scratch pad of notes and questions I collect as I read more about it.
Background
I’m coming from the context of building RAG applications that use vector embeddings to do semantic search across a set of documents. Familiarity with generating queries to the vector db and injecting the responses in the prompts for more context aware answers.
From here, I’ve kept hearing interest in using Graphs and GraphDBs instead of vector DBs with projects linked langraph or mindgraph. I’ve also seen the article GraphRAG referenced several times and, also keep seeing the Neo4j database pop up in my day-to-day life. So I decided to read more into Graphs to see what the hype was about.
GraphRAG
Using existing text data the language model is deciding what the entities and relationships are and generate Neo4j cypher code.
That is used for generating the graphs. Now how does that work with knowledge graphs?
Are graphs more efficient represents or encodings of information? If so are they a better language for language models to encode data and feed context and in that sense more helpful for them to use.
It kind of makes sense a conversation about having the nodes but also embeddeding the cosine similarity of everything within the graph.
Would be interesting to translate that all to a graph database. Can build a layer on top of collections.
Also fundamentally if I had access to a vector embedding space to alist of random facts, it would be harder to get information out of it than a structured graph with more metadata and information and intermediate reasoning about the structure of the data. Essentially a knowledge graph made by the agent would include more thoughts and reasoning about the user.
Other interesting readings I found on this rabbit hole were.
Remaining Questions
- How does the embedding dimensions map to the size of a document and in general
what is the max size of a document?
- In other terms how do I know how much information is too much for one document
- How hard is it to create graph representations of users.
- For code understanding how would you encode an Abstract Syntax Tree as a graph available to the language model?
- Is the langchain method of just chunking each individual section enough?
- Probably not because you can’t get relationships captured well with just a vector embedding.
- Also worth noting the LangChain method uses Tree-sitter
- How can you do a hybrid approach of combining vector embeddings with graphs
and also encoding the relationships of cosine similarity into the graph.
- Is there something similar to Matryoshka Embeddings in Graphs.
- I wonder if I can use Postgres with GraphDBs and stick to the use postgres for everything mindset. See these two relevant posts