Class: LangsmithrbRails::Evals::Checks::LlmGraded

Inherits:
Object
  • Object
show all
Defined in:
lib/generators/langsmithrb_rails/evals/templates/checks/llm_graded.rb

Overview

LLM-based grading for evaluating responses

Class Method Summary collapse

Class Method Details

.call_llm(prompt) ⇒ String

Call the LLM for grading



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
 
# File 'lib/generators/langsmithrb_rails/evals/templates/checks/llm_graded.rb', line 82

def self.call_llm(prompt)
  # Check if OpenAI is configured
  if defined?(OpenAI) && ENV["OPENAI_API_KEY"].present?
    client = OpenAI::Client.new(access_token: ENV["OPENAI_API_KEY"])
    response = client.chat(
      parameters: {
        model: ENV.fetch("LANGSMITH_EVAL_MODEL", "gpt-3.5-turbo"),
        messages: [{ role: "user", content: prompt }],
        temperature: 0.0
      }
    )
    return response.dig("choices", 0, "message", "content")
  end
  
  # Check if Anthropic is configured
  if defined?(Anthropic) && ENV["ANTHROPIC_API_KEY"].present?
    client = Anthropic::Client.new(api_key: ENV["ANTHROPIC_API_KEY"])
    response = client.messages.create(
      model: ENV.fetch("LANGSMITH_EVAL_MODEL", "claude-2"),
      max_tokens: 1024,
      messages: [{ role: "user", content: prompt }]
    )
    return response.content.first.text
  end
  
  # Fall back to a simple evaluation
  "Score: 0.5\nReasoning: Unable to perform LLM-based evaluation. Please configure an LLM provider."
end

.create_grading_prompt(input, answer, expected_answer) ⇒ String

Create a prompt for the LLM to grade the response



57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
 
# File 'lib/generators/langsmithrb_rails/evals/templates/checks/llm_graded.rb', line 57

def self.create_grading_prompt(input, answer, expected_answer)
  "    You are an expert evaluator. Your task is to grade the quality and correctness of a response.\n    \n    Question: \#{input[\"question\"]}\n    \n    Expected Answer: \#{expected_answer}\n    \n    Actual Response: \#{answer}\n    \n    Please evaluate the response based on:\n    1. Correctness: Is the information accurate?\n    2. Completeness: Does it fully address the question?\n    3. Clarity: Is it well-explained and easy to understand?\n    \n    Provide your evaluation in the following format:\n    \n    Score: [a number between 0.0 and 1.0]\n    Reasoning: [your detailed explanation]\n  PROMPT\nend\n"

.evaluate(input, response, expected) ⇒ Hash

Check if the response is correct using an LLM



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
 
# File 'lib/generators/langsmithrb_rails/evals/templates/checks/llm_graded.rb', line 13

def self.evaluate(input, response, expected)
  result = {
    score: 0.0,
    reasoning: "",
    passed: false
  }
  
  # Extract the answer from the response
  answer = extract_answer(response)
  expected_answer = extract_answer(expected)
  
  # Create the prompt for the LLM
  prompt = create_grading_prompt(input, answer, expected_answer)
  
  # Call the LLM for grading
  llm_response = call_llm(prompt)
  
  # Parse the LLM response
  parsed_result = parse_llm_response(llm_response)
  
  # Update the result with the parsed data
  result[:score] = parsed_result[:score]
  result[:reasoning] = parsed_result[:reasoning]
  result[:passed] = parsed_result[:score] >= 0.7
  
  result
end

.extract_answer(response) ⇒ String

Extract the answer from the response



44
45
46
47
48
49
50
 
# File 'lib/generators/langsmithrb_rails/evals/templates/checks/llm_graded.rb', line 44

def self.extract_answer(response)
  return response["answer"] if response["answer"]
  return response["text"] if response["text"]
  return response["content"] if response["content"]
  return response["output"] if response["output"]
  return response.to_s
end

.parse_llm_response(response) ⇒ Hash

Parse the LLM response



114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
 
# File 'lib/generators/langsmithrb_rails/evals/templates/checks/llm_graded.rb', line 114

def self.parse_llm_response(response)
  result = {
    score: 0.5,
    reasoning: "Unable to parse LLM response"
  }
  
  # Extract score
  if response =~ /Score:\s*([\d\.]+)/i
    result[:score] = $1.to_f
    # Ensure score is between 0 and 1
    result[:score] = [0.0, [1.0, result[:score]].min].max
  end
  
  # Extract reasoning
  if response =~ /Reasoning:\s*(.+)/im
    result[:reasoning] = $1.strip
  end
  
  result
end