models/hf_boss_api.py

import os
import io
import base64
import requests
import re

def hf_boss_answer(image, question, lang="en"):
    token = os.getenv("HF_TOKEN")
    if not token:
        return "Setup Required", "HF Token is missing.", "Please set HF_TOKEN."

    # Using Qwen 2.5 VL as requested
    model_id = "Qwen/Qwen2.5-VL-7B-Instruct:hyperbolic"
    
    try:
        api_url = "https://router.huggingface.co/v1/chat/completions"
        headers = {
            "Authorization": f"Bearer {token}",
            "Content-Type": "application/json"
        }
        
        if image.mode != 'RGB':
            image = image.convert('RGB')
        
        # Consistent resizing for vision models
        image.thumbnail((1024, 1024))
        buffered = io.BytesIO()
        image.save(buffered, format="JPEG", quality=85)
        img_b64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
        img_data_url = f"data:image/jpeg;base64,{img_b64}"

        prompt = f"""Analyze this image and answer the question.
        
Question: {question}

CRITICAL FORMATTING RULES:
1. All output must be in language: {lang}.
2. You MUST use exactly these THREE labels: "Caption:", "Final Answer:", and "Explanation:".
3. Do not use any markdown formatting like **bold** in the labels.
4. "Caption:" must be 1 line.
5. "Final Answer:" must be 1 line.
6. "Explanation:" must be 2-3 sentences.

Output Format:
Caption: [1 line]
Final Answer: [1 line]
Explanation: [2-3 sentences]"""

        payload = {
            "model": model_id,
            "messages": [
                {
                    "role": "system",
                    "content": "You are a professional VQA assistant. You follow formatting instructions perfectly."
                },
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": prompt},
                        {"type": "image_url", "image_url": {"url": img_data_url}}
                    ]
                }
            ],
            "max_tokens": 1024,
            "temperature": 0.1
        }
        
        response = requests.post(api_url, headers=headers, json=payload, timeout=90)
        
        # Fallback system
        if response.status_code != 200:
            model_id = "google/gemma-3-12b-it:featherless"
            payload["model"] = model_id
            response = requests.post(api_url, headers=headers, json=payload, timeout=90)

        if response.status_code != 200:
            model_id = "zai-org/GLM-4.5V:novita"
            payload["model"] = model_id
            response = requests.post(api_url, headers=headers, json=payload, timeout=90)

        if response.status_code != 200:
            return "HF Boss API Error", f"HTTP {response.status_code}", response.text
            
        data = response.json()
        raw_text = data["choices"][0]["message"]["content"]
        
        # AGGRESSIVE EXTRACTION LOGIC
        # This regex handles cases with or without bolding, and different spacings
        def extract(label, text):
            # Look for **Label**, Label:, etc.
            pattern = rf'(?:\*\*|#)?{label}(?:\*\*|#)?\s*[:\-]?\s*(.*?)(?=\s*(?:\n(?:\*\*|#)?[A-Z][a-z]+ [A-Z][a-z]+:?|\n(?:\*\*|#)?[A-Z][a-z]+:?)|$)'
            match = re.search(pattern, text, re.IGNORECASE | re.DOTALL)
            return match.group(1).strip() if match else None

        caption = extract("Caption", raw_text)
        answer = extract("Final Answer", raw_text)
        explanation = extract("Explanation", raw_text)

        # Emergency Fallback: If regex fails, use the raw text split by lines
        if not caption or not answer:
            lines = [l.strip() for l in raw_text.split('\n') if l.strip()]
            caption = caption or (lines[0] if len(lines) > 0 else "Analysis failed to parse.")
            answer = answer or (lines[1] if len(lines) > 1 else "No answer found in output.")
            explanation = explanation or (" ".join(lines[2:]) if len(lines) > 2 else raw_text)

        # Cleanup
        caption = caption.replace("**", "").replace("\n", " ").strip()
        answer = answer.replace("**", "").replace("\n", " ").strip()
        explanation = explanation.replace("**", "").strip()

        return caption, answer, explanation
        
    except Exception as e:
        return "HF Boss Internal Crash", "Parsing Error", f"Raw: {locals().get('raw_text', 'No text')} | Error: {repr(e)}"