Multi-doc graph merging seeding hallucination

[azman-nd]

When I upload multiple CV/resume docs, lightrag merges the entities/relationships of multiple CVs, when possible, into a graph. This is giving rise to pathway in graph that in really do not exists in resumes and are factually incorrect.
For example, Say Candidate A has done Project M using Angular (a programming language). Candidate B has done Project N with Angular also. Now after the graph is created both Candidate A and B is connected to "Angular" node. and Both Project M and N is also connected to Angular node.
In this case, If I retrieve, "Angular" node and it's surrounding node and give it to LLM, then LLM sees following paths

1. Candidate A-->Angular--> Project M
2. Candidate A--> Angular --> Project N
3. Candidate B-->Angular-->Project M
4. Candidate B-->Angular-->Project N

Now, obviously, #1 and #4 are factually correct but #2 and #3 are factually incorrect. These incorrect pathways seeds hallucination in LLMs.

How to prevent this graph contamination that arise from merging new doc content into existing graph?
Appreciate your thoughts.

[pranav-tandon]

Potential Solutions

1. Document-Scoped Entity Names
   One approach would be to scope entity names by document during extraction, creating document-specific entities like "Angular_Doc1" and "Angular_Doc2" that only get merged if they're truly the same entity across documents.

2. Relationship Provenance Filtering
   Modify the query process to filter relationships based on document provenance. The system already tracks chunk IDs in entity_chunks_storage operate.py:1620-1634 , so queries could be enhanced to only return relationships that co-occur in the same document chunks.