Global search answers broad, thematic questions by using precomputed summaries of graph communities instead of retrieving individual text chunks through vector similarity. It converts fragmented information into a cohesive response across the dataset. section 7.3.1 section 7.0
Implement the retrieval flow as a map-reduce pipeline:
- Map the user query over each selected community report, producing intermediate key points with numerical importance ratings. section 7.3.1
- Reduce the intermediate results by filtering and aggregating the most important points. section 7.3.1
- Give the refined points to an LLM to synthesize the final answer. section 7.3.1
Choose the community level deliberately. Lower-level reports preserve more detail but require more LLM calls and processing time; higher-level reports improve efficiency while risking lost granularity. section 7.3.1
Story-level question
Calling global_retriever("What is this story about?") applies the strategy to a broad thematic query, aggregating evidence across community summaries rather than focusing on one entity. section 7.3.1
