All files / api/src/connectors/evaluations/argument-correctness/templates main.ts

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
 
 
 
 
 
1x
2x
2x
2x
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2x
 
 
2x
 
 
2x
 
 
2x
 
 
 
2x
2x
2x
2x
2x
2x
 
import type {
  ArgumentCorrectnessTemplateConfig,
  ArgumentCorrectnessTemplateData,
} from '@api/connectors/evaluations/argument-correctness/types';
 
export function getArgumentCorrectnessTemplate(
  data: ArgumentCorrectnessTemplateData,
): ArgumentCorrectnessTemplateConfig {
  const systemPrompt = `You are an expert evaluator for agentic systems. Assess whether the arguments provided to each tool call are correct given the task.
 
Instructions:
- Analyze the user input, the tools called (name, description, and input), and the agent's output.
- For each tool call, decide if the arguments are correct based on the task.
- Compute an overall score as (# of correct tool calls) / (total number of tool calls).
 
Return a JSON object with fields:
{
  "score": <number between 0 and 1>,
  "reasoning": "<concise explanation of overall assessment>",
  "metadata": {
    "per_tool": [ { "name": "<tool name>", "correct": <true|false>, "explanation": "<brief why>" } ]
  }
}`;
 
  const userPrompt = `Evaluate the following:
 
Input:
${data.input ?? ''}
 
Agent Output:
${data.actual_output ?? ''}
 
Tools Called (name, description, input):
${JSON.stringify(data.tools_called ?? [], null, 2)}
 
Judge whether the arguments for each tool are correct given the input.`;
 
  return {
    systemPrompt,
    userPrompt,
    outputFormat: 'json',
  } as const;
}