Class: Evals::PromptEvaluator

Inherits:
Object
  • Object
show all
Defined in:
lib/evals/prompt_evaluator.rb

Instance Method Summary collapse

Constructor Details

#initialize(max_concurrent_tasks: 3) ⇒ PromptEvaluator

Returns a new instance of PromptEvaluator.



7
8
9
10
11
 
# File 'lib/evals/prompt_evaluator.rb', line 7

def initialize(max_concurrent_tasks: 3)
  @max_concurrent_tasks = max_concurrent_tasks
  @client = Anthropic::Client.new(api_key: ENV["ANTHROPIC_API_KEY"])
  @model = "claude-3-5-haiku-latest"
end

Instance Method Details

#add_assistant_message(messages, text) ⇒ Object 



30
31
32
 
# File 'lib/evals/prompt_evaluator.rb', line 30

def add_assistant_message(messages, text)
  messages << {role: "assistant", content: text}
end

#add_user_message(messages, text) ⇒ Object 



26
27
28
 
# File 'lib/evals/prompt_evaluator.rb', line 26

def add_user_message(messages, text)
  messages << {role: "user", content: text}
end

#chat(messages, system: nil, temperature: 1.0, stop_sequences: []) ⇒ Object 



34
35
36
37
38
39
40
41
42
43
44
45
46
47
 
# File 'lib/evals/prompt_evaluator.rb', line 34

def chat(messages, system: nil, temperature: 1.0, stop_sequences: [])
  params = {
    model: @model,
    max_tokens: 1000,
    messages: messages,
    temperature: temperature,
    stop_sequences: stop_sequences
  }

  params[:system] = system if system

  response = @client.messages.create(params)
  response.content[0].text
end

#generate_dataset(task_description, prompt_inputs_spec: {}, num_cases: 1, output_file: "dataset.json") ⇒ Object 



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
 
# File 'lib/evals/prompt_evaluator.rb', line 195

def generate_dataset(task_description, prompt_inputs_spec: {}, num_cases: 1, output_file: "dataset.json")
  ideas = generate_unique_ideas(task_description, prompt_inputs_spec, num_cases)

  dataset = []
  completed = 0
  total = ideas.length
  last_reported_percentage = 0

  threads = ideas.map do |idea|
    Thread.new do
      generate_test_case(task_description, idea, prompt_inputs_spec)
    end
  end

  threads.each do |thread|
    result = thread.value
    completed += 1
    current_percentage = ((completed.to_f / total) * 100).to_i
    milestone_percentage = (current_percentage / 20) * 20

    if milestone_percentage > last_reported_percentage
      puts "Generated #{completed}/#{total} test cases"
      last_reported_percentage = milestone_percentage
    end

    dataset << result
  rescue => e
    puts "Error generating test case: #{e}"
  end

  File.write(output_file, JSON.pretty_generate(dataset))
  dataset
end

#generate_test_case(task_description, idea, prompt_inputs_spec = {}) ⇒ Object 



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
 
# File 'lib/evals/prompt_evaluator.rb', line 117

def generate_test_case(task_description, idea, prompt_inputs_spec = {})
  example_prompt_inputs = ""
  prompt_inputs_spec.each do |key, value|
    val = value.gsub("\n", "\\n")
    example_prompt_inputs += "\"#{key}\": \"EXAMPLE_VALUE\", // #{val}\n"
  end

  allowed_keys = prompt_inputs_spec.keys.map { |key| "\"#{key}\"" }.join(", ")

  prompt = "    Generate a single detailed test case for a prompt evaluation based on:\n\n    <task_description>\n    \#{task_description}\n    </task_description>\n\n    <specific_idea>\n    \#{idea}\n    </specific_idea>\n\n    <allowed_input_keys>\n    \#{allowed_keys}\n    </allowed_input_keys>\n\n    Output Format:\n    ```json\n    {\n        \"prompt_inputs\": {\n        \#{example_prompt_inputs}\n        },\n        \"solution_criteria\": [\"criterion 1\", \"criterion 2\", ...] // Concise list of criteria for evaluating the solution, 1 to 4 items\n    }\n    ```\n\n    IMPORTANT REQUIREMENTS:\n    - You MUST ONLY use these exact input keys in your prompt_inputs: \#{allowed_keys}\n    - Do NOT add any additional keys to prompt_inputs\n    - All keys listed in allowed_input_keys must be included in your response\n    - Make the test case realistic and practically useful\n    - Include measurable, concise solution criteria\n    - The solution criteria should ONLY address the direct requirements of the task description and the generated prompt_inputs\n    - Avoid over-specifying criteria with requirements that go beyond the core task\n    - Keep solution criteria simple, focused, and directly tied to the fundamental task\n    - The test case should be tailored to the specific idea provided\n    - Quick to solve without requiring extensive computation or multi-step processing\n    - Solvable with no more than 400 tokens of output\n    - DO NOT include any fields beyond those specified in the output format\n  TEXT\n\n  system_prompt = \"You are a test case creator specializing in designing evaluation scenarios.\"\n\n  rendered_prompt = render(\n    prompt.strip,\n    {\n      \"allowed_keys\" => allowed_keys,\n      \"task_description\" => task_description,\n      \"idea\" => idea,\n      \"example_prompt_inputs\" => example_prompt_inputs\n    }\n  )\n\n  messages = []\n  add_user_message(messages, rendered_prompt)\n  add_assistant_message(messages, \"```json\")\n  text = chat(\n    messages,\n    stop_sequences: [\"```\"],\n    system: system_prompt,\n    temperature: 0.7\n  )\n\n  test_case = JSON.parse(text)\n  test_case[\"task_description\"] = task_description\n  test_case[\"scenario\"] = idea\n\n  test_case\nend\n"

#generate_unique_ideas(task_description, prompt_inputs_spec, num_cases) ⇒ Object 



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
 
# File 'lib/evals/prompt_evaluator.rb', line 49

def generate_unique_ideas(task_description, prompt_inputs_spec, num_cases)
  prompt = "    Generate \#{num_cases} unique, diverse ideas for testing a prompt that accomplishes this task:\n\n    <task_description>\n    \#{task_description}\n    </task_description>\n\n    The prompt will receive the following inputs\n    <prompt_inputs>\n    \#{prompt_inputs_spec}\n    </prompt_inputs>\n\n    Each idea should represent a distinct scenario or example that tests different aspects of the task.\n\n    Output Format:\n    Provide your response as a structured JSON array where each item is a brief description of the idea.\n\n    Example:\n    ```json\n    [\n        \"Testing with technical computer science terminology\",\n        \"Testing with medical research findings\",\n        \"Testing with complex mathematical concepts\",\n        ...\n    ]\n    ```\n\n    Ensure each idea is:\n    - Clearly distinct from the others\n    - Relevant to the task description\n    - Specific enough to guide generation of a full test case\n    - Quick to solve without requiring extensive computation or multi-step processing\n    - Solvable with no more than 400 tokens of output\n\n    Remember, only generate \#{num_cases} unique ideas\n  TEXT\n\n  system_prompt = \"You are a test scenario designer specialized in creating diverse, unique testing scenarios.\"\n\n  example_prompt_inputs = \"\"\n  prompt_inputs_spec.each do |key, value|\n    val = value.gsub(\"\\n\", \"\\\\n\")\n    example_prompt_inputs += \"\\\"\#{key}\\\": str # \#{val},\"\n  end\n\n  rendered_prompt = render(\n    prompt.strip,\n    {\n      \"task_description\" => task_description,\n      \"num_cases\" => num_cases,\n      \"prompt_inputs\" => example_prompt_inputs\n    }\n  )\n\n  messages = []\n  add_user_message(messages, rendered_prompt)\n  add_assistant_message(messages, \"```json\")\n  text = chat(\n    messages,\n    stop_sequences: [\"```\"],\n    system: system_prompt,\n    temperature: 1.0\n  )\n\n  JSON.parse(text)\nend\n"

#grade_output(test_case, output, extra_criteria) ⇒ Object 



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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
 
# File 'lib/evals/prompt_evaluator.rb', line 229

def grade_output(test_case, output, extra_criteria)
  prompt_inputs = ""
  test_case["prompt_inputs"].each do |key, value|
    val = value.gsub("\n", "\\n")
    prompt_inputs += "\"#{key}\":\"#{val}\",\n"
  end

  extra_criteria_section = ""
  if extra_criteria
    extra_criteria_template = "      Mandatory Requirements - ANY VIOLATION MEANS AUTOMATIC FAILURE (score of 3 or lower):\n      <extra_important_criteria>\n      \#{extra_criteria}\n      </extra_important_criteria>\n    TEXT\n    extra_criteria_section = render(\n      extra_criteria_template.strip,\n      {\"extra_criteria\" => extra_criteria}\n    )\n  end\n\n  eval_template = <<~TEXT\n    Your task is to evaluate the following AI-generated solution with EXTREME RIGOR.\n\n    Original task description:\n    <task_description>\n    \#{test_case[\"task_description\"]}\n    </task_description>\n\n    Original task inputs:\n    <task_inputs>\n    { \#{prompt_inputs} }\n    </task_inputs>\n\n    Solution to Evaluate:\n    <solution>\n    \#{output}\n    </solution>\n\n    Criteria you should use to evaluate the solution:\n    <criteria>\n    \#{test_case[\"solution_criteria\"].join(\"\\n\")}\n    </criteria>\n\n    \#{extra_criteria_section}\n\n    Scoring Guidelines:\n    * Score 1-3: Solution fails to meet one or more MANDATORY requirements\n    * Score 4-6: Solution meets all mandatory requirements but has significant deficiencies in secondary criteria\n    * Score 7-8: Solution meets all mandatory requirements and most secondary criteria, with minor issues\n    * Score 9-10: Solution meets all mandatory and secondary criteria\n\n    IMPORTANT SCORING INSTRUCTIONS:\n    * Grade the output based ONLY on the listed criteria. Do not add your own extra requirements.\n    * If a solution meets all of the mandatory and secondary criteria give it a 10\n    * Don't complain that the solution \"only\" meets the mandatory and secondary criteria. Solutions shouldn't go above and beyond - they should meet the exact listed criteria.\n    * ANY violation of a mandatory requirement MUST result in a score of 3 or lower\n    * The full 1-10 scale should be utilized - don't hesitate to give low scores when warranted\n\n    Output Format\n    Provide your evaluation as a structured JSON object with the following fields, in this specific order:\n    - \"strengths\": An array of 1-3 key strengths\n    - \"weaknesses\": An array of 1-3 key areas for improvement\n    - \"reasoning\": A concise explanation of your overall assessment\n    - \"score\": A number between 1-10\n\n    Respond with JSON. Keep your response concise and direct.\n  TEXT\n\n  eval_prompt = render(\n    eval_template.strip,\n    {\n      \"task_description\" => test_case[\"task_description\"],\n      \"prompt_inputs\" => prompt_inputs,\n      \"output\" => output,\n      \"solution_criteria\" => test_case[\"solution_criteria\"].join(\"\\n\"),\n      \"extra_criteria_section\" => extra_criteria_section\n    }\n  )\n\n  messages = []\n  add_user_message(messages, eval_prompt)\n  add_assistant_message(messages, \"```json\")\n  eval_text = chat(\n    messages,\n    stop_sequences: [\"```\"],\n    temperature: 0.0\n  )\n\n  JSON.parse(eval_text)\nend\n"

#render(template_string, variables) ⇒ Object 



13
14
15
16
17
18
19
20
21
22
23
24
 
# File 'lib/evals/prompt_evaluator.rb', line 13

def render(template_string, variables)
  placeholders = template_string.scan(/{([^{}]+)}/)

  result = template_string
  placeholders.flatten.each do |placeholder|
    if variables.key?(placeholder)
      result = result.gsub("{#{placeholder}}", variables[placeholder].to_s)
    end
  end

  result.gsub("{{", "{").gsub("}}", "}")
end

#run_evaluation(run_prompt_function, dataset_file, extra_criteria: nil, json_output_file: "output.json", html_output_file: "output.html") ⇒ Object 



336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
 
# File 'lib/evals/prompt_evaluator.rb', line 336

def run_evaluation(run_prompt_function, dataset_file, extra_criteria: nil, json_output_file: "output.json", html_output_file: "output.html")
  dataset = JSON.parse(File.read(dataset_file))

  results = []
  completed = 0
  total = dataset.length
  last_reported_percentage = 0

  threads = dataset.map do |test_case|
    Thread.new do
      run_test_case(test_case, run_prompt_function, extra_criteria)
    end
  end

  threads.each do |thread|
    result = thread.value
    completed += 1
    current_percentage = ((completed.to_f / total) * 100).to_i
    milestone_percentage = (current_percentage / 20) * 20

    if milestone_percentage > last_reported_percentage
      puts "Graded #{completed}/#{total} test cases"
      last_reported_percentage = milestone_percentage
    end

    results << result
  end

  average_score = results.sum { |result| result["score"] } / results.length.to_f
  puts "Average score: #{average_score}"

  File.write(json_output_file, JSON.pretty_generate(results))

  html = generate_prompt_evaluation_report(results)
  File.write(html_output_file, html)

  results
end

#run_test_case(test_case, run_prompt_function, extra_criteria = nil) ⇒ Object 



321
322
323
324
325
326
327
328
329
330
331
332
333
334
 
# File 'lib/evals/prompt_evaluator.rb', line 321

def run_test_case(test_case, run_prompt_function, extra_criteria = nil)
  output = run_prompt_function.call(test_case["prompt_inputs"])

  model_grade = grade_output(test_case, output, extra_criteria)
  model_score = model_grade["score"]
  reasoning = model_grade["reasoning"]

  {
    "output" => output,
    "test_case" => test_case,
    "score" => model_score,
    "reasoning" => reasoning
  }
end