Current

Graph recall and corpus aggregation

Benchmark first

Corpora: DocRED · WDC Products · KILT · QASC

Protocol: Measure blocking pairs completeness and reduction ratio, entity pair/cluster F1, relation micro-F1, provenance precision/recall, and multi-hop answer accuracy. Then run graph-off and graph-on retrieval with identical seeds to attribute recall gains. Every projected passage must retain its source node and path.

Passage retrieval and corpus summarization are different operations. Personalized PageRank is a bounded multi-hop recall function. Deterministic community partitioning and model-injected map-reduce reports are separately versioned aggregation stages.

How it works

Link query mentions to authorized seed nodes, induce an allowed subgraph, and propagate Personalized PageRank mass until tolerance or iteration limits; project node mass back to evidence-bearing sections. Separately, Leiden partitions the graph into well-connected communities, recursive grouping forms levels, and evidence-linked community reports support global map-reduce queries. Local queries fan out from entities; drift queries start globally and open bounded local branches.

Papers

HippoRAG: Personalized PageRank recall Leiden community detection GraphRAG: community reports and global query

Authorized recall and community map-reduce

from mari_components.graph import (
    build_community_reports,
    leiden_communities,
    map_reduce_reports,
)
from mari_components.retrieval import personalized_pagerank

recall = personalized_pagerank(
    graph,
    seeds=query_seeds,
    allowed_node_ids=authorized_node_ids,
    damping=0.50,
)

partition = leiden_communities(
    graph,
    resolution=1.0,
    allowed_node_ids=authorized_node_ids,
)
reports = build_community_reports(partition, summarize=summarize_nodes)
answer = map_reduce_reports(
    reports,
    map_report=lambda report: answer_from_report(query, report),
    reduce_answers=lambda partials: synthesize(query, partials),
    limit=24,
)