forked from hesabix/arc
98 lines
No EOL
2 KiB
JSON
98 lines
No EOL
2 KiB
JSON
{
|
|
"dataset_version": "2026-08-19.v1",
|
|
"index": {
|
|
"kind": "inverted_keyword_v1",
|
|
"tools": 180,
|
|
"postings": 522,
|
|
"build_ms": 15.352,
|
|
"memory_bytes_est": 120048
|
|
},
|
|
"candidate_recall": [
|
|
{
|
|
"n": 20,
|
|
"queries": 443,
|
|
"candidate_recall": 0.9458,
|
|
"latency_p50_ms": 0.758,
|
|
"latency_p95_ms": 0.999,
|
|
"avg_latency_ms": 0.757
|
|
},
|
|
{
|
|
"n": 50,
|
|
"queries": 443,
|
|
"candidate_recall": 0.9865,
|
|
"latency_p50_ms": 0.841,
|
|
"latency_p95_ms": 1.137,
|
|
"avg_latency_ms": 0.826
|
|
},
|
|
{
|
|
"n": 100,
|
|
"queries": 443,
|
|
"candidate_recall": 0.9887,
|
|
"latency_p50_ms": 0.848,
|
|
"latency_p95_ms": 1.338,
|
|
"avg_latency_ms": 0.881
|
|
},
|
|
{
|
|
"n": 200,
|
|
"queries": 443,
|
|
"candidate_recall": 0.9887,
|
|
"latency_p50_ms": 0.858,
|
|
"latency_p95_ms": 1.347,
|
|
"avg_latency_ms": 0.892
|
|
}
|
|
],
|
|
"chosen_candidate_n": 100,
|
|
"target_candidate_recall": 0.99,
|
|
"indexed_final": [
|
|
{
|
|
"k": 10,
|
|
"candidate_n": 100,
|
|
"recall": 0.9503,
|
|
"mrr": 0.7831,
|
|
"ndcg": 0.8012
|
|
},
|
|
{
|
|
"k": 15,
|
|
"candidate_n": 100,
|
|
"recall": 0.9707,
|
|
"mrr": 0.7845,
|
|
"ndcg": 0.8086
|
|
},
|
|
{
|
|
"k": 20,
|
|
"candidate_n": 100,
|
|
"recall": 0.9774,
|
|
"mrr": 0.7849,
|
|
"ndcg": 0.811
|
|
}
|
|
],
|
|
"baseline_final": [
|
|
{
|
|
"k": 10,
|
|
"recall": 0.9549
|
|
},
|
|
{
|
|
"k": 15,
|
|
"recall": 0.9752
|
|
},
|
|
{
|
|
"k": 20,
|
|
"recall": 0.9819
|
|
}
|
|
],
|
|
"no_match": {
|
|
"queries": 30,
|
|
"empty_candidate_set": 15,
|
|
"empty_rate": 0.5
|
|
},
|
|
"flags": {
|
|
"INDEXED_CANDIDATE_RETRIEVAL": false,
|
|
"VECTOR_CANDIDATE_RETRIEVAL": false,
|
|
"CANDIDATE_RETRIEVAL_N": 100,
|
|
"HYBRID_TOOL_DISCOVERY": false,
|
|
"ADAPTIVE_DISCOVERY_K": false,
|
|
"ADAPTIVE_K_ROLLOUT_PERCENT": 0,
|
|
"PROGRESSIVE_SCHEMA_LOADING": false
|
|
},
|
|
"note": "Candidate retrieval is opt-in. Production still ranks the full authorized set at K=48/128. Vector ANN is not production."
|
|
} |