How are the edge weights computed when adding new documents?

[natoverse]

Copied from #630

Question text

Excuse me, may I ask how the weights of the triplets extracted from new documents are determined in the graph when adding multiple new documents? As stated in the paper, the weight of edges is the normalized count value. So, is it a reconstruction of all documents?

Answer

GraphRAG doesn't yet have an official "incremental" indexing strategy. However, you can achieve incremental indexing by adding new content and then re-running. This is not a commitment to a full re-run, because we aggressively cache responses as long as the parameters/config have not changed. Therefore, you can add documents and get an index update, which results in questions such as this.

Some parts of GraphRAG will naturally work with the cache because they are independent/atomic operations. Others will need to re-run any time the inputs have changed. So, when adding new documents you can expect any existing text units to remain, and all entities + relationships in those text units will also be pulled from the cache. However, in the graph construction step, the entire graph will be recreated because there will be new entities and relationships (nodes and edges) from the new content. Therefore, the edge weights will be recomputed across the entire graph.