feat: add regression tests and configure its workflow

This commit is contained in:
Lucas Gomide
2025-07-15 18:45:26 -03:00
parent 6ebb6c9b63
commit cbe570088e
7 changed files with 343 additions and 3 deletions

View File

@@ -0,0 +1,42 @@
[
{
"timestamp": "2025-07-15T21:34:08.253410+00:00",
"metadata": {},
"results": [
{
"identifier": "72239c22b0cdde98ad5c588074ef6325",
"inputs": {
"company": "Apple Inc. (AAPL)"
},
"score": {
"goal_alignment": 10.0,
"semantic_quality": 9.0,
"tool_selection": 6.0,
"parameter_extraction": 5.0,
"tool_invocation": 10.0,
"reasoning_efficiency": 7.300000000000001
},
"expected_score": {
"goal_alignment": 8
},
"passed": true
},
{
"identifier": "test_2",
"inputs": {
"company": "Microsoft Corporation (MSFT)"
},
"score": {
"goal_alignment": 10.0,
"semantic_quality": 7.333333333333333,
"tool_selection": 6.25,
"parameter_extraction": 9.5,
"tool_invocation": 10.0,
"reasoning_efficiency": 6.0
},
"expected_score": 8,
"passed": true
}
]
}
]

View File

@@ -0,0 +1,24 @@
[
{
"timestamp": "2025-07-15T21:31:05.916161+00:00",
"metadata": {},
"results": [
{
"identifier": "df0ea31ac4a7fb4a908b8319ec7b3719",
"inputs": {
"messages": "How was the Battle of Waterloo?"
},
"score": {
"goal_alignment": 10.0,
"semantic_quality": 10.0,
"tool_selection": 10.0,
"parameter_extraction": 10.0,
"tool_invocation": 10.0,
"reasoning_efficiency": 5.5
},
"expected_score": 8,
"passed": true
}
]
}
]