Эксперимент: Seed-Anchored Budgeted Graph Context (#1210)

{ "worked": true, "confidence": 9, "verdict": "Built a deterministic seed-matched BFS graph-context renderer with hop-tier ordering, stable identifiers, exact separator-inclusive lengths, and greedy budget truncation. The formal check showed an exact transition at D_k(S)=B=4119 characters: all tested budgets below D were incomplete, while B=D and larger budgets achieved candidate recall 1.0; repeated anchored renders were identical. Across 8 synthetic queries whose budgets fit all seed-local one-hop evidence, anchored rendering achieved 1.0 average relevant recall versus 0.273 for global ordering and 0.316 for random truncation.", "metrics": { "baseline": "Global ID ordering: average relevant recall 0.273 across 8 queries; random ordering: 0.316.", "idea": "Seed-anchored hop-tier ordering: average relevant recall 1.000, with full relevant recall on 8/8 queries; formal candidate recall transitioned exactly at D_k(S)=B=4119 characters." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py && /home/maxwelhelp/main/bin/python3 mini_experiment.py", "files": [ "experiment.py", "mini_experiment.py", "results.json", "mini_results.json" ], "limitations": "Only synthetic graphs and exact name matching were tested; no graph-RAG reader, QA accuracy, learned retriever, pretrained model, tokenization effects, latency, or real-world noisy queries were evaluated. The evidence-recall comparison used budgets deliberately sized to fit the one-hop seed-local region, so it verifies retrieval behavior rather than end-to-end reader gains." }