I haven't really seen this, but the best way to start is creating a simple test case that reproduces some unexpected results 🙂 from there, it's 100x easier to debug
Do you know what exactly is getting embedded in the property graph? Is it the entire node? or just the name? or node and some of the paths?
damn, thanks callam, this is what I was looking
it should be just the entity name right now
@Logan M Okay and just for sanity that only node/entity that this isn't true for is the chunk nodes correct?
@Logan M how Do i obtain the embedings of the named entities retrieved?
We have been doing it through Neo4J
We can't seem to get the vector retriever working properly
We are in Google Meet trying to figure it out rn
It seems like there is a bug in the Vector retriever we have doubled checked the embeddings in both the query and the node
Also would like to know if possible to add an edge based of a condition of the embedings and if so, how
yep, add customized edges to nodes based on the embeddings, that is
I still don't know what you mean
Can you give me an example?
like why do you wanna do this?
sometimes entities relationship are not catched, based on similarity scores I would like to be able to link them, just that
There is an article on entity resolution
Its like a collab between the Neo4J team and llama Index
@Logan M Idk if it was a bug but the issue we were finding is that the get_rel_map function had a query that didn't consider the order of the nodes being passed in so it would just grab nodes that could be totally unrelated and it would fill the limit with them before ever getting to the most relevant nodes. We are optimizing that right now
@Logan M I wanted to ask is there a reason that we search the entities (not chunks) from the vector search? My thought is that this will break often when you are searching for stuff you didn't explicitly mention in your property graph. Seems like it should get the Chunks that are relevant then grab all the entities that we have mentioned. Just wanted to see if there was a reason for doing it entity first. Instead of Chunk first?
Would you be down to jump on a call at somepoint?
It would be good to chat?
Chunk first would just be a normal vector db no? Instead, we can search entity first, and map back to chunk. Happy to have a retriever that does that opposite though
Yeah I was thinking it would be a normal Vector DB first but then give the context of the entities a chunk is attached to?
So like If a chunk is talking to about a person, and organization and project we would get all the information about the person, organization and project that we know through out the database instead of just what is talked about in that chunk. I think we will make that retriever also cause we need it!
the mapping function is not implemented yet, just use the nodes
