import os
import json
import tqdm
from openai import OpenAI

# --- CONFIGURATION ---
MODEL_PATH = "/home/mshahidul/readctrl_model/full_model/qwen3-32B_subclaims-support-check-8b_ctx_v2-bf16"
API_URL = "http://172.16.34.29:8004/v1" 
API_KEY = "EMPTY"

client = OpenAI(base_url=API_URL, api_key=API_KEY)

def get_reasoning_prompt_json(source_text, gold_summary, generated_text, subclaim, level):
    """
    Forces the model to output a machine-readable JSON object for clinical logic validation.
    """
    return f"""You are a clinical logic validator auditing medical text simplification.

### Context & Goals:
- **Target Literacy Level:** {level}

1. Level: Low Health Literacy (High Readability)

Target: Individuals needing the simplest terms for immediate action.

Linguistic Goal: Use "living room" language. Replace all medical jargon with functional descriptions (e.g., "renal" becomes "kidney").

Information Density: Focus strictly on the "need-to-know" info found in the Gold Summary.

Strategy: High paraphrasing using analogies. One idea per sentence.

Faithfulness: Must align perfectly with the Gold Summary.

2. Level: Intermediate Health Literacy (Medium Readability)

Target: The general public (news-reading level).

Linguistic Goal: Standard vocabulary. Common medical terms are okay, but technical "doctor-speak" must be simplified.

Information Density: Balanced. Use the Gold Summary as the lead, supplemented by necessary context from the Source Text.

Strategy: Moderate paraphrasing. Remove minor technical details to avoid information overload.

Faithfulness: Maintains the main narrative of the Gold Summary.

3. Level: Proficient Health Literacy (Low Readability)

Target: Researchers, clinicians, or highly informed patients.

Linguistic Goal: Technical and academic language. Prioritize clinical nuance and medical accuracy.

Information Density: High. Use the Full Source Text to include data, physiological mechanisms, and statistics.

Strategy: Minimal paraphrasing. Retain all original technical terminology.

Faithfulness: Adhere to the Source Text; you may add related subclaims that provide deeper scientific context.

### Input Data:
1. **Source Text:** {source_text}
2. **Gold Summary (Reference):** {gold_summary}
3. **Generated Text (Output):** {generated_text}
4. **Subclaim to Evaluate:** {subclaim}

### Task:
Evaluate the Subclaim's status in the Generated Text compared to the Source and Gold Summary. Output ONLY a JSON object.

### Classification Categories:
- "reasonable_removal": Subclaim in Source, but NOT in Gold (non-essential).
- "reasonable_modification": Subclaim simplified correctly for the {level} goal.
- "unreasonable_removal": Subclaim in Gold but MISSING from Generated (critical loss).
- "unreasonable_addition": Subclaim in Generated but NOT in Source/Gold (hallucination).
- "preserved": Fact maintained with high fidelity.

### JSON Schema Requirement:
{{
  "category": "string (reasonable_removal | reasonable_modification | unreasonable_removal | unreasonable_addition | preserved)",
  "action": "string (added | removed | modified | preserved)",
  "presence_in_gold": "boolean",
  "presence_in_generated": "boolean",
  "verdict": "string (one sentence clinical justification)"
}}

Output JSON:"""

def evaluate_reasoning_json(source, gold, generated, subclaim, level):
    prompt = get_reasoning_prompt_json(source, gold, generated, subclaim, level)
    
    try:
        response = client.chat.completions.create(
            model=MODEL_PATH,
            messages=[{"role": "user", "content": prompt}],
            max_tokens=400,
            temperature=0.1,
            # If your vLLM setup supports JSON mode, you can add response_format={"type": "json_object"}
        )
        content = response.choices[0].message.content.strip()
        
        # Clean potential markdown formatting if model outputs ```json ... ```
        if content.startswith("```json"):
            content = content.replace("```json", "").replace("```", "").strip()
        
        return json.loads(content)
    except Exception as e:
        return {
            "category": "error",
            "action": "error",
            "verdict": f"API or Parsing Error: {str(e)}"
        }

# -----------------------------
# Example Usage in your Main Loop:
# -----------------------------
# result = evaluate_reasoning_json(full_text, ref_summary, summary_at_level, sc, level)
# print(result['category'])