1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
|
#!/usr/bin/env ruby
# frozen_string_literal: true
require 'gitlab'
require 'json'
class Reporter
GITLAB_COM_API_V4_ENDPOINT = "https://gitlab.com/api/v4"
QA_EVALUATION_PROJECT_ID = 52020045 # https://gitlab.com/gitlab-org/ai-powered/ai-framework/qa-evaluation
AGGREGATED_REPORT_ISSUE_IID = 1 # https://gitlab.com/gitlab-org/ai-powered/ai-framework/qa-evaluation/-/issues/1
IDENTIFIABLE_NOTE_TAG = 'gitlab-org/ai-powered/ai-framework:duo-chat-qa-evaluation'
GRADE_TO_EMOJI_MAPPING = {
correct: ":white_check_mark:",
incorrect: ":x:",
unexpected: ":warning:"
}.freeze
def run
if pipeline_running_on_master_branch?
snippet_web_url = upload_data_as_snippet
report_issue_url = create_report_issue
update_aggregation_issue(report_issue_url, snippet_web_url)
else
save_report_as_artifact
post_or_update_report_note
end
end
def markdown_report
@report ||= <<~MARKDOWN
<!-- #{IDENTIFIABLE_NOTE_TAG} -->
## GitLab Duo Chat QA evaluation
Report generated for "#{ENV['CI_JOB_NAME']}". This report is generated and refreshed automatically. Do not edit.
LLMs have been asked to evaluate GitLab Duo Chat's answers.
:white_check_mark: : LLM evaluated the answer as `CORRECT`.
:x: : LLM evaluated the answer as `INCORRECT`.
:warning: : LLM did not evaluate correctly or the evaluation request might have failed.
### Summary
- The total number of evaluations: #{summary_numbers[:total]}
- The number of evaluations in which all LLMs graded `CORRECT`: #{summary_numbers[:correct]} (#{summary_numbers[:correct_ratio]}%)
- Note: if an evaluation request failed or its response was not parsable, it was ignored. For example, :white_check_mark: :warning: would count as `CORRECT`.
- The number of evaluations in which all LLMs graded `INCORRECT`: #{summary_numbers[:incorrect]} (#{summary_numbers[:incorrect_ratio]}%)
- Note: if an evaluation request failed or its response was not parsable, it was ignored. For example, :x: :warning: would count as `INCORRECT`.
- The number of evaluations in which LLMs disagreed: #{summary_numbers[:disagreed]} (#{summary_numbers[:disagreed_ratio]}%)
### Evaluations
#{eval_content}
MARKDOWN
# Do this to avoid pinging users in notes/issues.
quote_usernames(@report)
end
private
def quote_usernames(text)
text.gsub(/(@\w+)/, '`\\1`')
end
def pipeline_running_on_master_branch?
ENV['CI_COMMIT_BRANCH'] == ENV['CI_DEFAULT_BRANCH']
end
def utc_timestamp
@utc_timestamp ||= Time.now.utc
end
def upload_data_as_snippet
filename = "#{utc_timestamp.to_i}.json"
title = utc_timestamp.to_s
snippet_content = ::JSON.pretty_generate({
commit: ENV["CI_COMMIT_SHA"],
pipeline_url: ENV["CI_PIPELINE_URL"],
data: report_data
})
puts "Creating a snippet #{filename}."
snippet = qa_evaluation_project_client.create_snippet(
QA_EVALUATION_PROJECT_ID,
{
title: title,
files: [{ file_path: filename, content: snippet_content }],
visibility: 'private'
}
)
snippet.web_url
end
def create_report_issue
puts "Creating a report issue."
issue_title = "Report #{utc_timestamp}"
new_issue = qa_evaluation_project_client.create_issue(
QA_EVALUATION_PROJECT_ID, issue_title, { description: markdown_report }
)
new_issue.web_url
end
def update_aggregation_issue(report_issue_url, snippet_web_url)
puts "Updating the aggregated report issue."
new_line = ["\n|"]
new_line << "#{utc_timestamp} |"
new_line << "#{summary_numbers[:total]} |"
new_line << "#{summary_numbers[:correct_ratio]}% |"
new_line << "#{summary_numbers[:incorrect_ratio]}% |"
new_line << "#{summary_numbers[:disagreed_ratio]}% |"
new_line << "#{report_issue_url} |"
new_line << "#{snippet_web_url} |"
new_line = new_line.join(' ')
aggregated_report_issue = qa_evaluation_project_client.issue(QA_EVALUATION_PROJECT_ID, AGGREGATED_REPORT_ISSUE_IID)
updated_description = aggregated_report_issue.description + new_line
qa_evaluation_project_client.edit_issue(
QA_EVALUATION_PROJECT_ID, AGGREGATED_REPORT_ISSUE_IID, { description: updated_description }
)
end
def save_report_as_artifact
artifact_path = File.join(base_dir, ENV['QA_EVAL_REPORT_FILENAME'])
puts "Saving #{artifact_path}"
File.write(artifact_path, markdown_report)
end
def post_or_update_report_note
note = existing_report_note
if note && note.type != 'DiscussionNote'
# The latest note has not led to a discussion. Update it.
gitlab_project_client.edit_merge_request_note(ci_project_id, merge_request_iid, note.id, markdown_report)
puts "Updated comment."
else
# This is the first note or the latest note has been discussed on the MR.
# Don't update, create new note instead.
gitlab_project_client.create_merge_request_note(ci_project_id, merge_request_iid, markdown_report)
puts "Posted comment."
end
end
def existing_report_note
# Look for an existing note using `IDENTIFIABLE_NOTE_TAG`
gitlab_project_client
.merge_request_notes(ci_project_id, merge_request_iid)
.auto_paginate
.select { |note| note.body.include? IDENTIFIABLE_NOTE_TAG }
.max_by { |note| Time.parse(note.created_at) }
end
def gitlab_project_client
@gitlab_project_client ||= Gitlab.client(
endpoint: GITLAB_COM_API_V4_ENDPOINT,
private_token: ENV['PROJECT_TOKEN_FOR_CI_SCRIPTS_API_USAGE']
)
end
def qa_evaluation_project_client
@qa_evaluation_project_client ||= Gitlab.client(
endpoint: GITLAB_COM_API_V4_ENDPOINT,
private_token: ENV['CHAT_QA_EVALUATION_PROJECT_TOKEN_FOR_CI_SCRIPTS_API_USAGE']
)
end
def base_dir
ENV['CI_PROJECT_DIR'] || "./"
end
def merge_request_iid
ENV['CI_MERGE_REQUEST_IID']
end
def ci_project_id
ENV['CI_PROJECT_ID']
end
def report_data
@report_data ||= Dir[File.join(base_dir, "tmp/duo_chat/qa*.json")]
.flat_map { |file| JSON.parse(File.read(file)) }
end
def eval_content
report_data
.sort_by { |a| a["question"] }
.map do |data|
<<~MARKDOWN
<details>
<summary>
#{correctness_indicator(data)}
`"#{data['question']}"`
(context: `#{data['resource']}`)
</summary>
#### Resource
`#{data['resource']}`
#### Answer
#{data['answer']}
#### LLM Evaluation
Tools used: #{data['tools_used']}
#{evalutions(data)}
</details>
MARKDOWN
end
.join
end
def summary_numbers
@graded_evaluations ||= report_data
.map { |data| data["evaluations"].map { |eval| parse_grade(eval) } }
.reject { |grades| !(grades.include? :correct) && !(grades.include? :incorrect) }
total = @graded_evaluations.size
correct = @graded_evaluations.count { |grades| !(grades.include? :incorrect) }
incorrect = @graded_evaluations.count { |grades| !(grades.include? :correct) }
disagreed = @graded_evaluations.count { |grades| (grades.include? :correct) && (grades.include? :incorrect) }
{
total: total,
correct: correct,
correct_ratio: (correct.to_f / total * 100).round(1),
incorrect: incorrect,
incorrect_ratio: (incorrect.to_f / total * 100).round(1),
disagreed: disagreed,
disagreed_ratio: (disagreed.to_f / total * 100).round(1)
}
end
def parse_grade(eval)
return :correct if eval["response"].match?(/Grade: CORRECT/i)
return :incorrect if eval["response"].match?(/Grade: INCORRECT/i)
# If the LLM's evaluation includes neither CORRECT nor CORRECT, flag it.
:unexpected
end
def correctness_indicator(data)
data["evaluations"].map { |eval| parse_grade(eval) }.map { |grade| GRADE_TO_EMOJI_MAPPING[grade] }.join(' ')
end
def evalutions(data)
rows = data["evaluations"].map do |eval|
grade = parse_grade(eval)
<<~MARKDOWN
<tr>
<td>#{eval['model']}</td>
<td>
#{GRADE_TO_EMOJI_MAPPING[grade]}
</td>
<td>
#{eval['response']}
</td
</tr>
MARKDOWN
end
.join
<<~MARKDOWN
<table>
<tr>
<td>Model</td>
<td>Grade</td>
<td>Details</td>
</tr>
#{rows}
</table>
MARKDOWN
end
end
Reporter.new.run if $PROGRAM_NAME == __FILE__
|