Context. Nautilus fans out to heterogeneous sources and returns an unranked {source_id: rows} passthrough. MemGraphRAG's central empirical finding: high-recall / low-relevance context degrades downstream quality — the same hazard for a broker handing an agent every row from every source. A ranking synthesizer models the recall-vs-relevance trade-off without touching the deterministic routing path. Depends on the merge-context widening (linked issue).
Scope note (paper-faithfulness). PPR / hub-suppression run over a pre-built heterogeneous KG in the paper; transplanting them onto an ad-hoc query-time source/row bipartite graph is an analogy, not a port. The defensible, paper-grounded core here is similarity-to-query reranking + top-K trim (optionally a batch-local IDF down-weight). PPR framing is explicitly out of scope.
Proposal. New nautilus/synthesis/ranking.py RankingSynthesizer.merge(results, context):
- Seed per-row scores by similarity to
context['embedding'] (or a per-source/broker Embedder), optional batch-local IDF down-weight on recurring key tokens, keep top-K.
- Pure-Python, no LLM, deterministic. Preserve the no-raise contract + error pre-filtering.
- Degrade to passthrough when no embedding available (default is
NoopEmbedder(strict=True), embedder.py:48).
Where.
nautilus/synthesis/base.py (Protocol), new nautilus/synthesis/ranking.py.- Select in
Broker.from_config (replaces hardcoded BasicSynthesizer() at broker.py:581); invoked broker.py:1380. Embedder Protocol adapters/embedder.py.
Acceptance criteria.
Hash caveat. Reordering rows changes state.data feeding response_hash via compute_response_hash(state.data) (broker.py:1950; list order IS hash-significant). Per-source source_response_hashes (broker.py:1664) are computed over RAW pre-synthesis rows and are unaffected.
Source: arXiv:2606.00610v1 — "MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation". Distilled from arXiv-research/2606.00610v1/analysis.md; file refs verified against current main by the analysis pass.
Context. Nautilus fans out to heterogeneous sources and returns an unranked
{source_id: rows}passthrough. MemGraphRAG's central empirical finding: high-recall / low-relevance context degrades downstream quality — the same hazard for a broker handing an agent every row from every source. A ranking synthesizer models the recall-vs-relevance trade-off without touching the deterministic routing path. Depends on the merge-context widening (linked issue).Scope note (paper-faithfulness). PPR / hub-suppression run over a pre-built heterogeneous KG in the paper; transplanting them onto an ad-hoc query-time source/row bipartite graph is an analogy, not a port. The defensible, paper-grounded core here is similarity-to-query reranking + top-K trim (optionally a batch-local IDF down-weight). PPR framing is explicitly out of scope.
Proposal. New
nautilus/synthesis/ranking.pyRankingSynthesizer.merge(results, context):context['embedding'](or a per-source/broker Embedder), optional batch-local IDF down-weight on recurring key tokens, keep top-K.NoopEmbedder(strict=True),embedder.py:48).Where.
nautilus/synthesis/base.py(Protocol), newnautilus/synthesis/ranking.py.Broker.from_config(replaces hardcodedBasicSynthesizer()atbroker.py:581); invokedbroker.py:1380. Embedder Protocoladapters/embedder.py.Acceptance criteria.
Hash caveat. Reordering rows changes
state.datafeedingresponse_hashviacompute_response_hash(state.data)(broker.py:1950; list order IS hash-significant). Per-sourcesource_response_hashes(broker.py:1664) are computed over RAW pre-synthesis rows and are unaffected.Source: arXiv:2606.00610v1 — "MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation". Distilled from
arXiv-research/2606.00610v1/analysis.md; file refs verified against currentmainby the analysis pass.