Reference

Graph recall and corpus aggregation¶

Behavior¶

Mechanism	Corpus observation	Decision guidance
Personalized PageRank	QASC passage Recall@5 `0.850`. Both gold facts found `0.709`	Graph expansion exposed many two-hop facts. Complete chains appeared in `0.709` of cases
Local-modularity community heuristic	DocRED relation-community coverage `0.430`. Mean modularity `0.518`	Co-mention graphs form coherent groups and split many gold relations
Community reports	Deterministic partition and bounded map/reduce	Report quality depends on the injected summarizer

Example graph propagation example

Node	Relation to seed	Allowed	Retrieval outcome
`refund-policy`	Seed	Yes	Ranked
`enterprise-exception`	Two hops away	Yes	Activated by propagation
`private-escalation`	One hop away	No	Removed before propagation

Community input	Output invariant
Disconnected candidate cluster	Split into connected components
Same graph and resolution	Same deterministic partition

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.

leiden_communities is a historical compatibility name for deterministic local modularity improvement followed by connected-component splitting. It averages opposing edge weights and drops self-links. Its output is a flat heuristic partition. Full Leiden additionally requires an aggregation phase. For native hierarchical Leiden, Louvain, and bounded DRIFT search, see the independently selectable algorithm choices. Evidence binding and model callbacks remain caller-supplied.

Papers

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

Authorized recall and community map-reduce¶

from mari_kit.graph import (
    build_community_reports,
    leiden_communities,
    map_reduce_reports,
)
from mari_kit.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,
)

build_community_reports invokes the summarizer once per community and retains node IDs with each report. map_reduce_reports processes the first limit reports in supplied order. Rank reports before this call when relevance should determine the subset. Token budgets and model-call execution belong to those callbacks.

Incremental report maintenance¶

Treat a community report as a derived artifact with explicit membership, source revisions, summarizer version, and configuration inputs. Community IDs such as community:0 are local partition positions. Assign durable scoped identity before caching reports across partitions. The shared dependency planner can reuse completed reports whose inputs match and hold dependent answers until changed reports finish rebuilding.