A gleaning loop improves entity and relationship extraction during graph indexing. After an initial LLM pass, it shows the model its own output and requests additional items that were missed, repeating for a fixed number of iterations. section 12.2.1
Recovering overlooked incident details
From a passage about Anna Müller, AuthService, and INC-2347, the first pass may extract entities and relationships; subsequent passes ask the model to add omitted details, improving recall in dense, entity-heavy text. section 12.2.1
The procedure follows a compact indexing pattern:
- Extract entities with types and descriptions, plus relationships with descriptions and numeric strengths. section 12.2.1
- Show the model its prior output and prompt it to add missed entities and relationships. section 12.2.1
- Bind results to source chunks so later graph nodes retain provenance. section 12.2.1
