Does Graph-Augmented Retrieval Help on a Small Regulatory Corpus? A Controlled Comparison with Whole-Corpus Prompting

On a 12-document aerospace corpus, a graph-augmented retrieval pipeline did not answer or attribute better than vector-only retrieval or than putting the whole corpus in the prompt. We compared four configurations on a 34-question golden set, 3 runs each, from a clean commit, with one shared citation rule: every answer cites documents in a fixed tag format, and citations are scored only on canonical document ids. The full pipeline had the highest golden pass rate (0.824 vs 0.667–0.745), but no paired difference was significant after Holm correction. It brought far more documents into context (context citation recall 0.985 vs 0.763–0.783, significant), yet its answers did not cite more accurately: answer citation F1 was 0.709, against 0.767 for vector-only. Whole-corpus prompting answered best on content (0.922 pass with citation checks excluded) and had the highest answer citation recall, but the lowest RAGAS faithfulness (0.803, significantly below the full pipeline). At this scale, the extra retrieval machinery widens the context without improving what the model writes.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24345271
Primary Topic
Topic Modeling
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Does Graph-Augmented Retrieval Help on a Small Regulatory Corpus? A Controlled Comparison with Whole-Corpus Prompting

Sergiu Nicoara
2 links
Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
article

Does Graph-Augmented Retrieval Help on a Small Regulatory Corpus? A Controlled Comparison with Whole-Corpus Prompting

Sergiu Nicoara
article en

Abstract

On a 12-document aerospace corpus, a graph-augmented retrieval pipeline did not answer or attribute better than vector-only retrieval or than putting the whole corpus in the prompt. We compared four configurations on a 34-question golden set, 3 runs each, from a clean commit, with one shared citation rule: every answer cites documents in a fixed tag format, and citations are scored only on canonical document ids. The full pipeline had the highest golden pass rate (0.824 vs 0.667–0.745), but no paired difference was significant after Holm correction. It brought far more documents into context (context citation recall 0.985 vs 0.763–0.783, significant), yet its answers did not cite more accurately: answer citation F1 was 0.709, against 0.767 for vector-only. Whole-corpus prompting answered best on content (0.922 pass with citation checks excluded) and had the highest answer citation recall, but the lowest RAGAS faithfulness (0.803, significantly below the full pipeline). At this scale, the extra retrieval machinery widens the context without improving what the model writes.

Zenodo (CERN European Organization for Nuclear Research)
Oldham Council (GB)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon