Backend - Checking on multiple LLMs for entity and relations extraction and populating KG

[aashipandya]

No description provided.

[jexp]

Comments from our discussion:

Thanks for putting the notebooks together my main feedback points summarized would be:

• have summary / conclusion for each in readme (also findings, learnings)
• use same PDF documents for each example
• generate the same machine processable (JSON) output from each (plus the human baseline) for comparison and further analysis
• use dotenv in the notebooks to allow loading environment variables from an .env file for the notebook so we don't need to ensure to add / remove credentials

Better use JSON than CSV actually because then we can handle multiple properties for each entity, align it with the structure we get from diffbot -> nodes / relationships

Just comments, no action needed here:

for the triple ones (like rebel / llama-index) our challenge here is that we can't use the results out of the box, we would either have to:

• modify them to output property graph nodes or relationships
• or post-process the triples to aggregate all entity attribute triples into properties and only keep the triples that represent semantic relationships as such
• or do this during insertion of the data into the graph - aggregating when inserting, e.g. initially create/merge the nodes with their ID and subsequently merge on id + add property and for the relationships find start and end node with label and id and create relationship

for the Rebel one in: def create_triplets(tx, triplet) : if we want to look at this approach in the future we should see if we can carry the entity-type over, so we can use not just the generic :Node but in addition also a label for the type like :Person or :Organization
and then also do the attribute aggregation there