Function call_evaluation_gemini

eval/src/eval_common.py:277–330

Evaluate a single image using Gemini API via OpenAI-compatible client.

(task)

# ============================================================

def call_evaluation_gemini(task):
    """Evaluate a single image using Gemini API via OpenAI-compatible client."""
    from openai import OpenAI

    index = task["index"]
    prompt = task["prompt"]
    testpoint = task["testpoint"]
    test_desc = task["test_desc"]
    img_path = task["img_path"]
    lang = task["lang"]
    max_retries = task["max_retries"]
    api_key = task["api_key"]
    base_url = task["base_url"]
    model_name = task.get("model_name", "gemini-2.5-pro")

    system_prompt = build_system_prompt(prompt, testpoint, test_desc, lang)
    base64_image = local_image_to_data_url(img_path)
    client = OpenAI(api_key=api_key, base_url=base_url)

    text = None
    for attempt in range(max_retries + 1):
        try:
            response = client.chat.completions.create(
                model=model_name,
                messages=[{
                    "role": "user",
                    "content": [
                        {"type": "text", "text": system_prompt},
                        {"type": "image_url", "image_url": {"url": base64_image}},
                    ],
                }],
                max_tokens=4096,
            )
            text = response.choices[0].message.content
        except Exception as e:
            print(f"Gemini API error (attempt {attempt + 1}/{max_retries + 1}): {e}")
            text = None

        if text is None:
            continue

        parsed = parse_evaluation_response(text, testpoint)
        if parsed is not None:
            analysis, score = parsed
            result_json = {
                "prompt": prompt,
                "testpoint": testpoint,
                "analysis": analysis,
                "score": score,
            }
            return make_result(index, testpoint, prompt, img_path, text, result_json)

    # All retries exhausted
    return make_result(index, testpoint, prompt, img_path, text)


# ============================================================