Current

Retrieval plans and context envelopes

Benchmark first

Corpora: BEIR (SciFact, ArguAna, HotpotQA) · LoTTE · KILT

Protocol: Build from corpus documents only; issue the published test queries; preserve ranked IDs and component scores. Report nDCG@10, Recall@100, MRR, p50/p95 query latency, index bytes, and results with ACL pre-filtering. Ablate candidate generation, reranking, fusion, graph expansion, and packing independently.

Mari packs already-ranked retrieval candidates into a bounded, revision-bearing context envelope. Its trace explains every inclusion and exclusion.

How it works

Run semantic, lexical, graph, and recency arms over authorized IDs. Convert arm scores to ranks and combine them with reciprocal-rank fusion, then discard stale dependencies, rerank survivors, diversify near-duplicates, and greedily pack whole evidence excerpts under token/document limits. The envelope contains rendered context plus source revisions and per-candidate include/exclude reasons, allowing the caller to reproduce what the model saw.

Research basisRAG motivates explicit, updateable non-parametric memory and provenance; RAG-Fusion and MMR back fusion and diversity; Lost in the Middle makes budget and evidence order evaluation requirements. ContextEnvelope is Mari’s carrier for those observable decisions.

semanticlexicalgraphrecent

RRF

authorizefreshnessrerank

budget

ContextEnvelopeexcerpts · evidence · revisions · trace

context.py

from mari_components.retrieval import ContextBudget, ContextCandidate, assemble_context

context = assemble_context([
    ContextCandidate(document_id=hit.document_id, revision=revisions[hit.document_id],
        text=passages[hit.document_id], token_count=token_count(hit.document_id),
        score=hit.score, authorized=can_read(hit.document_id),
        fresh=is_fresh(hit.document_id))
    for hit in fused_hits
], budget=ContextBudget(tokens=6000, documents=12))

model(context.text)
audit(context.trace)