ATLAS
Exploring the geometry of knowledge. Testing where structure helps—and where it doesn’t.
Can graph structure or persistent topology improve multi-hop retrieval beyond semantic similarity?
INTERACTIVE WORKFLOW
ATLAS / test the structural hypothesis
The experimental route adds graph and topology features. The selected V2 cosine-only configuration bypasses those stages after validation found no replicated benefit.
Dashed connection: selected cosine baseline bypasses experimental reranking. Solid connection: experimental comparison path.
Query
Evaluate a question against a fixed corpus; shared-support components separate internal splits.
Question + corpus
Read the complete workflow
- Query
Evaluate a question against a fixed corpus; shared-support components separate internal splits.
Documented experiment · Output: Question + corpus - Embedding
Encode query and documents into a shared semantic representation for cosine comparison.
Documented method · Output: Embedding vectors - Baseline retrieval
Exact cosine similarity produces the baseline ranking. The selected V2 path goes directly to evaluation.
Selected primary path · Output: Cosine ranked candidates - Graph / topology
Compare graph, topology-only and combined experimental reranking against the unchanged baseline.
Experimental / bypassed in primary · Output: Experimental rankings - Evaluation
Measure all-support@5 on held-out HotpotQA and MuSiQue subsets using shared evaluation rules.
Repository-reported evidence · Output: Saved rankings + metrics - Results
Cosine scores 0.5100 on HotpotQA versus 0.4700 for graph + topology. No replicated confirmatory improvement was established.
Not independently reproduced · Output: Negative finding + uncertainty
MEASURED EVIDENCE / REPOSITORY-REPORTED
The simpler baseline held up.
Held-out all-support@5: the fraction of questions with every gold supporting document in the first five retrieved results. Higher is better.
ALL-SUPPORT@5 / FIXED 0–1 SCALE
Repository-reported point estimates. Bars do not establish statistical significance; see the linked report for uncertainty intervals. No replicated improvement was established.
| Method | HotpotQA · n=400 | MuSiQue · n=300 |
|---|---|---|
| Cosine | 0.5100 | 0.1767 |
| Graph | 0.4875 | 0.1600 |
| Topology only | 0.4900 | 0.1800 |
| Graph + topology | 0.4700 | 0.1467 |
No replicated confirmatory improvement was established. The small MuSiQue topology-only point difference is not an established advantage.
Public development-data subsets, not official hidden-test or fullwiki scores. No answer generation was evaluated. Measurements have not been independently rerun for this portfolio.
Methods, uncertainty intervals & raw evidence ↗What the study tests
Graph structure and persistent topology might improve multi-hop all-support retrieval over cosine similarity. The reported comparisons did not establish that improvement.
Methodology and algorithms
A reproducible research pipeline compares semantic, graph, topology-only and combined retrieval. Separate local TIP and S-MCP prototypes explore region matching and constrained compute-to-data.
Controls and comparisons
The V2 configuration keeps the simpler cosine-only path after validation. Losing configurations and shuffled-feature controls remain part of the research record.
What the results say
The repository reports HotpotQA all-support@5 of 0.5100 for cosine and 0.4700 for graph + topology (400 held-out questions). The tested additions did not establish a replicated improvement. These reported measurements have not been independently rerun for this portfolio.
Where the evidence ends
Internal held-out subsets of public development data; no fullwiki or official hidden-test claim. Local prototypes establish neither privacy guarantees nor production security.
Tests and validation
V2 tests exercise retrieval and evaluation logic. The report retains cosine, graph-only, topology-only and combined results on HotpotQA and MuSiQue internal held-out subsets.
Reproduce the investigation
Follow docs/REPRODUCING.md and research/v2/PROTOCOL.md. Preserve dataset splits, feature configuration, shuffled-feature controls and evaluation rules; compare the committed report before interpreting new runs.
Source reviewed 17 October 2026. Summaries describe repository evidence; they do not assert sole authorship or independent reproduction.
Read the source documentation ↗Credit belongs to the repository’s contributors. Account or organization ownership does not establish sole authorship; individual contribution boundaries have not been verified.