aiqa-client 0.5.2__py3-none-any.whl → 0.6.1__py3-none-any.whl

aiqa/llm_as_judge.py

@@ -0,0 +1,281 @@
+"""
+LLM-as-judge scoring functionality for evaluating outputs.
+"""
+
+import os
+import json
+import re
+import asyncio
+import requests
+from typing import Any, Dict, Optional, Callable, Awaitable
+from .types import MetricResult, Metric, Example, CallLLMType
+
+def parse_llm_response(content: str) -> Optional[MetricResult]:
+    """
+    Parse LLM response content string into score and message.
+    
+    Args:
+        content: Raw content string from LLM response
+        
+    Returns:
+        MetricResult object with score:[0,1], message (optional), and error (optional)
+    """
+    try:
+        result = json.loads(content)
+        score = result.get("score")
+        if not score:
+            return None
+        message = result.get("message")
+        # Ensure score is in [0, 1] range
+        score = max(0.0, min(1.0, float(score)))
+        return {"score": score, "message": message}
+    except (json.JSONDecodeError, ValueError, TypeError) as e:
+        print(f"Failed to parse JSON response: {e}")
+        # Fallback: try to extract score from text
+        raise Exception(f"Could not parse LLM response: {content}")
+
+
+async def get_model_from_server(
+    model_id: str, server_url: str, headers: Dict[str, str]
+) -> Optional[Dict[str, Any]]:
+    """
+    Fetch a model from the server with API key included.
+    
+    Args:
+        model_id: ID of the model to fetch
+        server_url: URL of the AIQA server
+        headers: HTTP headers for authentication
+            
+    Returns:
+        Model dictionary with api_key if available, None if not found or error
+    """
+    try:
+        def _do_request():
+            return requests.get(
+                f"{server_url}/model/{model_id}?fields=api_key",
+                headers=headers,
+            )
+        
+        response = await asyncio.to_thread(_do_request)
+        if response.ok:
+            model = response.json()
+            if model.get("api_key"):
+                return model
+        return None
+    except Exception as e:
+        print(f"Warning: Could not fetch model from server: {e}")
+        return None
+
+
+async def call_openai(
+    system_prompt: str, user_message: str, api_key: str, output_schema: str = None
+) -> str:
+    """Call OpenAI API for LLM-as-judge scoring. Returns raw content string."""
+    def _do_request():
+        # TODO send output_schema if provided
+        return requests.post(
+            "https://api.openai.com/v1/chat/completions",
+            headers={
+                "Content-Type": "application/json",
+                "Authorization": f"Bearer {api_key}",
+            },
+            json={
+                "model": "gpt-4o-mini",  # Default model, can be made configurable
+                "messages": [
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": user_message},
+                ],
+                "temperature": 0,
+                "response_format": {"type": "json_object"},
+            },
+        )
+
+    response = await asyncio.to_thread(_do_request)
+
+    if not response.ok:
+        error_text = response.text
+        raise Exception(
+            f"OpenAI API error: {response.status_code} {response.reason} - {error_text}"
+        )
+
+    data = response.json()
+    choices = data.get("choices", [])
+    if not choices or not isinstance(choices, list):
+        raise Exception("OpenAI API did not return choices")
+    content = choices[0].get("message", {}).get("content")
+    if not content:
+        raise Exception("OpenAI API did not return content")
+
+    print(f"OpenAI raw response: {content[:500]}...")  # Show first 500 chars
+    return content
+
+
+async def call_anthropic(
+    system_prompt: str, user_message: str, api_key: str
+) -> str:
+    """Call Anthropic (Claude) API for LLM-as-judge scoring. Returns raw content string."""
+    def _do_request():
+        return requests.post(
+            "https://api.anthropic.com/v1/messages",
+            headers={
+                "Content-Type": "application/json",
+                "x-api-key": api_key,
+                "anthropic-version": "2023-06-01",
+            },
+            json={
+                "model": "claude-3-5-sonnet-20241022",  # Default model
+                "max_tokens": 1024,
+                "temperature": 0,
+                "system": system_prompt,
+                "messages": [{"role": "user", "content": user_message}],
+            },
+        )
+
+    response = await asyncio.to_thread(_do_request)
+
+    if not response.ok:
+        error_text = response.text
+        raise Exception(
+            f"Anthropic API error: {response.status_code} {response.reason} - {error_text}"
+        )
+
+    data = response.json()
+    content_list = data.get("content", [])
+    if not content_list or not isinstance(content_list, list):
+        raise Exception("Anthropic API did not return content")
+    content = content_list[0].get("text", "")
+    if not content:
+        raise Exception("Anthropic API did not return content")
+
+    print(f"Anthropic raw response: {content[:500]}...")  # Show first 500 chars
+    return content
+
+
+async def call_llm_fallback(
+    system_prompt: str,
+    user_message: str,
+    api_key: Optional[str] = None,
+    provider: Optional[str] = None,
+) -> str:
+    """
+    Fallback LLM call function that checks for API key parameter or environment variables.
+
+    Args:
+        system_prompt: System prompt for the LLM
+        user_message: User message containing the output to score
+        api_key: Optional API key to use (takes precedence over env vars)
+        provider: Optional provider name ('openai', 'anthropic', etc.) to determine which API to call
+
+    Returns:
+        Dictionary with "score" (float 0-1) and "message" (str)
+    """
+    # If API key provided, use it with the specified provider
+    if api_key:
+        if provider == 'openai' or provider is None:
+            content = await call_openai(system_prompt, user_message, api_key)
+        elif provider == 'anthropic':
+            content = await call_anthropic(system_prompt, user_message, api_key)
+        else:
+            # Try OpenAI first, then Anthropic for unknown providers
+            try:
+                content = await call_openai(system_prompt, user_message, api_key)
+            except Exception:
+                content = await call_anthropic(system_prompt, user_message, api_key)
+    else:
+        # Fallback to environment variables
+        openai_key = os.getenv("OPENAI_API_KEY")
+        anthropic_key = os.getenv("ANTHROPIC_API_KEY")
+
+        if openai_key:
+            print("Using OpenAI API (from OPENAI_API_KEY env var)")
+            content = await call_openai(system_prompt, user_message, openai_key)
+        elif anthropic_key:
+            print("Using Anthropic API (from ANTHROPIC_API_KEY env var)")
+            content = await call_anthropic(system_prompt, user_message, anthropic_key)
+        else:
+            raise Exception(
+                "No LLM API key found. Either:\n"
+                "  - Specify a model in metric.parameters.model (to use server-stored API key), or\n"
+                "  - Set OPENAI_API_KEY or ANTHROPIC_API_KEY environment variable (for local API key)"
+            )
+    return content
+
+
+async def score_llm_metric_local(
+    input_data: Any,
+    output: Any,
+    example: Example,
+    metric: Metric,
+    llm_call_fn: Optional[CallLLMType] = None,
+) -> MetricResult:
+    """
+    Score an LLM-as-judge metric.
+
+    Args:
+        input_data: The input data to score
+        output: The output to score
+        example: The example object
+        metric: The metric definition
+        llm_call_fn: Optional async function that takes (system_prompt, user_message) and returns
+                    raw content string (typically JSON). If not provided, will use fallback.
+
+    Returns:
+        MetricResult object with score:[0,1], message (optional), and error (optional)
+    """
+    # Build system prompt from metric description or prompt field
+    metric_prompt = metric.get("prompt")
+    if metric_prompt:
+        system_prompt = metric_prompt
+    else:
+        metric_text = metric.get("description", "") or metric.get("name", "")
+        system_prompt = f"""You are a judge evaluating AI assistant OUTPUTs for this metric:
+{metric_text}
+
+You MUST respond with ONLY a valid JSON object (no other text before or after) containing:
+- "score": a number between 0 and 1 (where 1 is best)
+- "message": a brief explanation of your scoring
+
+Example response format:
+{{"score": 0.75, "message": "The response was helpful but could be more concise"}}
+
+Be strict but fair in your evaluation."""
+
+    if isinstance(input_data, dict):
+        input_text = json.dumps(input_data)
+    else:
+        input_text = str(input_data)
+
+    if isinstance(output, dict):
+        output_text = json.dumps(output)
+    else:
+        output_text = str(output)
+
+    # Build user message with the input and output to score 
+    user_message = f"""<INPUT>:
+{input_text}
+</INPUT>
+<OUTPUT>
+{output_text}
+</OUTPUT>
+Evaluate this OUTPUT according to the metric and return ONLY a valid JSON object {{"score": number, "message": string}}."""
+
+    # Debug: print what we're sending to the LLM
+    print(f"DEBUG: Sending to LLM scorer:")
+    print(f"  System prompt: {system_prompt[:200]}...")
+    print(f"  User message: {user_message[:500]}...")
+    print(f"  Example input extracted: {repr(input_text[:100])}")
+
+    # Call LLM
+    print(f"Calling LLM to score metric '{metric.get('name')}'...")
+    if llm_call_fn:
+        result = await llm_call_fn(system_prompt, user_message)
+    else:
+        # Note: api_key and provider should be handled by the caller if model is specified
+        # This fallback uses environment variables
+        result = await call_llm_fallback(system_prompt, user_message)
+    
+    score_result = parse_llm_response(result)
+    if not score_result:
+        raise Exception(f"Failed to parse LLM response for metric {metric.get('id', 'unknown')}: {result}")
+    return score_result
+