Skip to content

5.4 Dataset Complexity Report

This generated report is the dataset-ladder and per-query view, not the canonical all-metric ranking presentation. The complete static base and flavor tables for every approach and metric are in evaluation-results.md.

It tracks approach rankings by input dataset, ordered from the simplest curated corpus to increasingly graph-heavy real-world candidates. It deliberately reports by dataset rather than by vector/graph collection, because the comparison question is how each RAG approach behaves as the input problem becomes more relational, temporal, and multi-hop. Each row also names the Atlas ingestion profile whose revision and job id are stored with newly generated matrix and judgment snapshots.

For the run protocol, model roles, approach invocation details, and judge-panel design, see evaluation-methodology.md. Committed evidence and provenance are listed in results/README.md. For approach-by-approach internals and tuning surfaces, see approaches.md.

1. Dataset Complexity Ladder

Dataset Complexity Status Atlas ingestion profile Graph nature Query file Source
baseline_curated 1 measured baseline_curated Mostly textual retrieval with a few multi-hop and exact-keyword prompts. demo/queries.yaml https://huggingface.co/datasets/yixuantt/MultiHopRAG
graph_native 2 measured graph_native Explicit entity and relationship bullets over AI, antitrust, crypto, regulators, witnesses, and timelines. demo/graph_native_queries.yaml corpus/graph_native
stark_prime 3 candidate stark_prime Biomedical entity retrieval over diseases, drugs, genes, pathways, proteins, phenotypes, and textual descriptions. demo/stark_prime_queries.yaml https://github.com/snap-stanford/stark
stark_mag 4 candidate stark_mag Paper, author, venue, field, citation, and affiliation retrieval where query constraints mix text with graph relations. demo/stark_mag_queries.yaml https://stark.stanford.edu/
openalex_scholarly 5 candidate openalex_scholarly Real scholarly graph with works, authors, institutions, concepts, citations, venues, and abstracts. demo/openalex_scholarly_queries.yaml https://developers.openalex.org/
gdelt_events 6 candidate gdelt_events Event, actor, location, theme, source, tone, and timeline graph over real news events. demo/gdelt_events_queries.yaml https://www.gdeltproject.org/
cyber_threat_intel 7 measured cyber_threat_intel Intrusion groups, campaigns, malware, tools, ATT&CK techniques, mitigations, and explicit uses/mitigates relationships. demo/cyber_threat_intel_queries.yaml https://attack.mitre.org/

2. Judge-Panel Ranking Drift by Input Dataset

Dataset Complexity Status Winner Ranking
baseline_curated 1 measured vanilla-rag vanilla-rag 4.17 > hybrid-rag 4.00 > contextual-rag 3.92 > lazy-graph-rag 3.92 > graph-rag 3.75 > n8n-adaptive-rag 3.33 > agentic-rag 2.67
graph_native 2 measured lazy-graph-rag lazy-graph-rag 4.31 > contextual-rag 4.19 > vanilla-rag 4.06 > hybrid-rag 3.62 > graph-rag 2.62 > agentic-rag 2.44 > n8n-adaptive-rag 2.44
stark_prime 3 candidate pending live run pending live run
stark_mag 4 candidate pending live run pending live run
openalex_scholarly 5 candidate pending live run pending live run
gdelt_events 6 candidate pending live run pending live run
cyber_threat_intel 7 measured contextual-rag contextual-rag 3.17 > agentic-rag 3.00 > lazy-graph-rag 3.00 > n8n-adaptive-rag 3.00 > vanilla-rag 3.00 > hybrid-rag 2.92 > graph-rag 2.42

3. Flavor-Tier Tuning Results

Flavor aliases are ranked separately from base families. They reuse the same dataset ingestion and graph state, so this tier measures query-time parameter choices without allowing several variants of one architecture to occupy the base-family leaderboard.

Dataset Status Flavor winner Flavor ranking
baseline_curated measured lazy-graph-rag-wide lazy-graph-rag-wide 4.58 > hybrid-rag-fast 4.08 > vanilla-rag-wide 4.08 > contextual-rag-high-recall 3.92 > graph-rag-wide 3.83 > lazy-graph-rag-fast 3.83 > hybrid-rag-high-recall 3.75 > lazy-graph-rag-balanced 3.75 > graph-rag-fast 3.67 > graph-rag-rerank 3.67 > n8n-adaptive-rag-default 3.17 > agentic-rag-deeper 2.92
graph_native measured hybrid-rag-high-recall hybrid-rag-high-recall 4.19 > vanilla-rag-wide 4.00 > lazy-graph-rag-balanced 3.69 > lazy-graph-rag-wide 3.69 > hybrid-rag-fast 3.62 > lazy-graph-rag-fast 3.62 > contextual-rag-high-recall 3.38 > graph-rag-wide 2.62 > graph-rag-rerank 2.44 > n8n-adaptive-rag-default 2.38 > agentic-rag-deeper 2.25 > graph-rag-fast 2.06
stark_prime candidate pending fresh flavor run pending fresh flavor run
stark_mag candidate pending fresh flavor run pending fresh flavor run
openalex_scholarly candidate pending fresh flavor run pending fresh flavor run
gdelt_events candidate pending fresh flavor run pending fresh flavor run
cyber_threat_intel measured hybrid-rag-fast hybrid-rag-fast 3.67 > contextual-rag-high-recall 3.50 > hybrid-rag-high-recall 3.33 > lazy-graph-rag-fast 3.33 > lazy-graph-rag-balanced 3.00 > vanilla-rag-wide 3.00 > lazy-graph-rag-wide 2.92 > n8n-adaptive-rag-default 2.58 > agentic-rag-deeper 2.42 > graph-rag-rerank 2.42 > graph-rag-wide 2.17 > graph-rag-fast 1.92

3.1 Flavor Per-Query Winners

Dataset Query Winner Top 3 mean scores
baseline_curated keyword graph-rag-wide agentic-rag-deeper 5.00 > contextual-rag-high-recall 5.00 > graph-rag-fast 5.00
baseline_curated thematic lazy-graph-rag-wide lazy-graph-rag-wide 5.00 > hybrid-rag-fast 4.50 > vanilla-rag-wide 4.50
baseline_curated multihop graph-rag-wide graph-rag-wide 4.50 > graph-rag-fast 3.00 > lazy-graph-rag-wide 3.00
baseline_curated factoid n8n-adaptive-rag-default contextual-rag-high-recall 5.00 > graph-rag-fast 5.00 > graph-rag-rerank 5.00
baseline_curated context_starved graph-rag-wide agentic-rag-deeper 5.00 > graph-rag-fast 5.00 > graph-rag-rerank 5.00
baseline_curated mixed_batch lazy-graph-rag-fast contextual-rag-high-recall 5.00 > hybrid-rag-fast 5.00 > hybrid-rag-high-recall 5.00
graph_native entity_bridge graph-rag-wide graph-rag-wide 5.00 > hybrid-rag-fast 5.00 > hybrid-rag-high-recall 5.00
graph_native relationship_chain vanilla-rag-wide contextual-rag-high-recall 5.00 > hybrid-rag-fast 5.00 > hybrid-rag-high-recall 5.00
graph_native shared_actor graph-rag-rerank graph-rag-rerank 4.50 > lazy-graph-rag-fast 3.50 > agentic-rag-deeper 3.00
graph_native timeline_cause lazy-graph-rag-wide lazy-graph-rag-wide 5.00 > contextual-rag-high-recall 4.50 > hybrid-rag-high-recall 4.50
graph_native witness_network vanilla-rag-wide hybrid-rag-high-recall 3.50 > lazy-graph-rag-balanced 3.50 > lazy-graph-rag-wide 3.50
graph_native cloud_model_competition hybrid-rag-high-recall hybrid-rag-high-recall 5.00 > agentic-rag-deeper 2.50 > contextual-rag-high-recall 2.50
graph_native default_search_ecosystem lazy-graph-rag-fast contextual-rag-high-recall 5.00 > hybrid-rag-high-recall 5.00 > lazy-graph-rag-balanced 5.00
graph_native cross_domain_regulators vanilla-rag-wide vanilla-rag-wide 5.00 > graph-rag-rerank 3.00 > hybrid-rag-fast 3.00
cyber_threat_intel cyber_group_technique_software_chain hybrid-rag-high-recall hybrid-rag-high-recall 4.50 > lazy-graph-rag-wide 4.50 > contextual-rag-high-recall 4.00
cyber_threat_intel cyber_credential_access_path hybrid-rag-fast hybrid-rag-fast 4.50 > contextual-rag-high-recall 4.00 > hybrid-rag-high-recall 4.00
cyber_threat_intel cyber_campaign_overlap hybrid-rag-fast hybrid-rag-fast 5.00 > contextual-rag-high-recall 4.00 > lazy-graph-rag-fast 3.50
cyber_threat_intel cyber_mitigation_coverage lazy-graph-rag-fast hybrid-rag-fast 3.50 > lazy-graph-rag-balanced 3.50 > lazy-graph-rag-fast 3.50
cyber_threat_intel cyber_campaign_timeline_context lazy-graph-rag-balanced lazy-graph-rag-balanced 4.00 > contextual-rag-high-recall 3.00 > hybrid-rag-fast 3.00
cyber_threat_intel cyber_protocol_and_web_mitigation_path vanilla-rag-wide vanilla-rag-wide 5.00 > hybrid-rag-high-recall 4.00 > lazy-graph-rag-fast 4.00

4. Canonical Evaluation Metrics

These columns come from the append-safe evidence rows. Ragas values are evaluator-model scores, latency and failures are operational measurements, and coverage is shown beside every ranking so unevaluable or failed cells cannot disappear from the comparison.

Dataset Faithfulness ranking Answer relevancy ranking Latency ranking Row coverage Failures
baseline_curated vanilla-rag 0.672 (6/6) > agentic-rag 0.664 (5/6) > hybrid-rag 0.645 (6/6) > lazy-graph-rag 0.636 (5/6) > n8n-adaptive-rag 0.564 (5/6) > contextual-rag 0.300 (5/6) contextual-rag 0.894 (6/6) > hybrid-rag 0.869 (6/6) > lazy-graph-rag 0.858 (6/6) > graph-rag 0.842 (6/6) > n8n-adaptive-rag 0.733 (6/6) > vanilla-rag 0.708 (6/6) > agentic-rag 0.567 (6/6) n8n-adaptive-rag 2798 ms (6/6) > vanilla-rag 3834 ms (6/6) > lazy-graph-rag 5514 ms (6/6) > agentic-rag 10774 ms (6/6) > graph-rag 12612 ms (6/6) > hybrid-rag 17683 ms (6/6) > contextual-rag 18016 ms (6/6) 42/42 successful 0 errors, 0 timeouts
graph_native contextual-rag 0.607 (5/8) > lazy-graph-rag 0.416 (5/8) > hybrid-rag 0.303 (6/8) > vanilla-rag 0.279 (5/8) > agentic-rag 0.000 (1/8) > n8n-adaptive-rag 0.000 (1/8) hybrid-rag 0.859 (8/8) > vanilla-rag 0.851 (8/8) > contextual-rag 0.841 (8/8) > lazy-graph-rag 0.829 (8/8) > graph-rag 0.796 (8/8) > agentic-rag 0.740 (8/8) > n8n-adaptive-rag 0.740 (8/8) vanilla-rag 4539 ms (8/8) > lazy-graph-rag 4936 ms (8/8) > n8n-adaptive-rag 5299 ms (8/8) > hybrid-rag 9809 ms (8/8) > contextual-rag 11197 ms (8/8) > graph-rag 12474 ms (8/8) > agentic-rag 29113 ms (8/8) 56/56 successful 0 errors, 0 timeouts
stark_prime not measured not measured not measured not available not available
stark_mag not measured not measured not measured not available not available
openalex_scholarly not measured not measured not measured not available not available
gdelt_events not measured not measured not measured not available not available
cyber_threat_intel contextual-rag 0.232 (2/6) > agentic-rag 0.224 (3/6) > n8n-adaptive-rag 0.224 (3/6) > lazy-graph-rag 0.121 (3/6) > vanilla-rag 0.052 (3/6) > hybrid-rag 0.000 (3/6) contextual-rag 0.891 (6/6) > hybrid-rag 0.878 (6/6) > agentic-rag 0.877 (6/6) > n8n-adaptive-rag 0.877 (6/6) > lazy-graph-rag 0.873 (6/6) > vanilla-rag 0.819 (6/6) > graph-rag 0.780 (6/6) lazy-graph-rag 8123 ms (6/6) > vanilla-rag 8907 ms (6/6) > n8n-adaptive-rag 11067 ms (6/6) > graph-rag 21203 ms (6/6) > hybrid-rag 22721 ms (6/6) > contextual-rag 24200 ms (6/6) > agentic-rag 41858 ms (6/6) 42/42 successful 0 errors, 0 timeouts

5. Base-Family Per-Query Winners

The Winner column is the judge panel's observed_winner: the approach with the highest mean score, breaking ties by best-answer votes. The Top 3 mean scores column ranks by mean only (ties ordered by name), so when several approaches tie on mean the vote-decided winner can fall outside the listed top three.

Dataset Query Winner Top 3 mean scores
baseline_curated keyword graph-rag agentic-rag 5.00 > contextual-rag 5.00 > graph-rag 5.00
baseline_curated thematic hybrid-rag contextual-rag 3.50 > hybrid-rag 3.50 > vanilla-rag 3.50
baseline_curated multihop graph-rag graph-rag 3.00 > agentic-rag 2.50 > n8n-adaptive-rag 2.50
baseline_curated factoid graph-rag contextual-rag 5.00 > graph-rag 5.00 > hybrid-rag 5.00
baseline_curated context_starved graph-rag agentic-rag 5.00 > graph-rag 5.00 > n8n-adaptive-rag 5.00
baseline_curated mixed_batch lazy-graph-rag lazy-graph-rag 5.00 > hybrid-rag 4.50 > vanilla-rag 4.50
graph_native entity_bridge vanilla-rag vanilla-rag 5.00 > agentic-rag 4.50 > n8n-adaptive-rag 4.50
graph_native relationship_chain vanilla-rag lazy-graph-rag 5.00 > vanilla-rag 5.00 > contextual-rag 4.50
graph_native shared_actor vanilla-rag vanilla-rag 4.50 > hybrid-rag 4.00 > lazy-graph-rag 4.00
graph_native timeline_cause hybrid-rag hybrid-rag 5.00 > contextual-rag 4.50 > lazy-graph-rag 4.50
graph_native witness_network contextual-rag contextual-rag 4.50 > lazy-graph-rag 4.50 > agentic-rag 3.00
graph_native cloud_model_competition contextual-rag contextual-rag 4.50 > hybrid-rag 3.50 > vanilla-rag 3.50
graph_native default_search_ecosystem lazy-graph-rag contextual-rag 5.00 > lazy-graph-rag 5.00 > vanilla-rag 4.00
graph_native cross_domain_regulators lazy-graph-rag lazy-graph-rag 4.50 > contextual-rag 3.00 > hybrid-rag 3.00
cyber_threat_intel cyber_group_technique_software_chain contextual-rag contextual-rag 5.00 > lazy-graph-rag 4.00 > vanilla-rag 3.50
cyber_threat_intel cyber_credential_access_path hybrid-rag hybrid-rag 5.00 > agentic-rag 3.00 > n8n-adaptive-rag 3.00
cyber_threat_intel cyber_campaign_overlap contextual-rag contextual-rag 4.50 > agentic-rag 4.00 > n8n-adaptive-rag 4.00
cyber_threat_intel cyber_mitigation_coverage vanilla-rag vanilla-rag 4.50 > lazy-graph-rag 3.50 > agentic-rag 2.50
cyber_threat_intel cyber_campaign_timeline_context lazy-graph-rag lazy-graph-rag 3.50 > contextual-rag 3.00 > agentic-rag 2.50
cyber_threat_intel cyber_protocol_and_web_mitigation_path hybrid-rag hybrid-rag 5.00 > graph-rag 3.50 > agentic-rag 3.00

6. Interpretation

The current measured ladder has 3 rungs. On baseline_curated, vanilla-rag leads; on graph_native, lazy-graph-rag leads; on cyber_threat_intel, contextual-rag leads. graph-rag is measured end to end across the live rungs but does not lead any of them.

That tells us the next step is not simply adding more documents; it is adding datasets whose native task requires relational retrieval, temporal event reasoning, and multi-hop graph paths.

The candidate rungs are intentionally heavier: STaRK-Prime and STaRK-MAG are semi-structured retrieval benchmarks; OpenAlex adds a real scholarly citation/author/institution graph; GDELT adds event-time actor/location graphs; and the measured cyber slice adds threat-technique, software, campaign, intrusion-group, and mitigation relationships. Scores for candidate rungs should be added only after live matrix and judge runs produce committed snapshots.

7. Candidate Dataset Sources

  • STaRK: semi-structured textual + relational retrieval benchmark with Amazon, MAG, and Prime domains.
  • OpenAlex: CC0 scholarly graph of works, authors, institutions, concepts, venues, and citations.
  • GDELT: global event/news graph with actors, events, locations, themes, sources, and timelines.
  • MITRE ATT&CK: measured bounded cyber graph over intrusion groups, campaigns, software, techniques, and mitigations.