forked from hesabix/arc
48 lines
No EOL
1,015 B
JSON
48 lines
No EOL
1,015 B
JSON
{
|
|
"k": 15,
|
|
"recall": 0.9752,
|
|
"multi_tool_recall": 0.9,
|
|
"mrr": 0.787,
|
|
"precision": 0.1706,
|
|
"ndcg": 0.8113,
|
|
"retrieval_success_rate": 0.9752,
|
|
"avg_schema_tokens": 3706,
|
|
"context_share": 0.1324,
|
|
"no_match_success": 0.6333,
|
|
"no_match_false_positive": 0.3667,
|
|
"no_match_false_negative": 0.0,
|
|
"per_category_recall": {
|
|
"CRM": 0.931,
|
|
"accounting": 0.9655,
|
|
"customer": 1.0,
|
|
"inventory": 1.0,
|
|
"invoice": 1.0,
|
|
"legacy": 1.0,
|
|
"memory": 1.0,
|
|
"purchase": 1.0,
|
|
"reports": 0.9474,
|
|
"sales": 1.0,
|
|
"support": 0.9259,
|
|
"wallet": 1.0,
|
|
"workflow": 1.0
|
|
},
|
|
"per_difficulty_recall": {
|
|
"ambiguous": 0.9565,
|
|
"easy": 0.9925,
|
|
"hard": 0.9551,
|
|
"medium": 0.9746
|
|
},
|
|
"per_hard_case_recall": {
|
|
"ambiguous": 1.0,
|
|
"conversational": 1.0,
|
|
"english": 0.9831,
|
|
"generic": 0.6667,
|
|
"long": 1.0,
|
|
"multi-intent": 1.0,
|
|
"persian_accounting": 1.0,
|
|
"short": 1.0,
|
|
"similar_tools": 1.0,
|
|
"synonyms": 1.0,
|
|
"typos": 1.0
|
|
}
|
|
} |