src/inference.py

import torch
import json
import numpy as np
from transformers import AutoModelForCausalLM, AutoTokenizer
from typing import Dict, List, Union
import re
import os

class ResumeInferenceEngine:
    """Inference engine for resume extraction and matching"""
    
    def __init__(self, model_path: str = "models/checkpoints/final"):
        """Load fine-tuned model and tokenizer"""
        
        print(f"Loading model from {model_path}...")

        # CPU-only environments (common on Windows laptops) can hit PEFT/accelerate
        # offload edge-cases when using device_map="auto". Prefer a simple CPU load.
        # RAM Optimization: Force half-precision (bfloat16 is best on CPU)
        # 16GB is tight for float32 loading, so we use bfloat16 to cut RAM in half.
        dtype = torch.bfloat16 if torch.cuda.is_available() or hasattr(torch, 'bfloat16') else torch.float32
        device_map = "auto" if torch.cuda.is_available() else None
        low_cpu_mem_usage = True # Always use low_cpu_mem_usage to prevent spikes

        adapter_config_path = os.path.join(model_path, "adapter_config.json")
        is_adapter = os.path.exists(adapter_config_path)

        # Prefer tokenizer saved alongside adapter/model (the notebook saves tokenizer to final/)
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
        if self.tokenizer.pad_token is None and self.tokenizer.eos_token is not None:
            self.tokenizer.pad_token = self.tokenizer.eos_token

        if is_adapter:
            from peft import PeftModel
            with open(adapter_config_path, "r", encoding="utf-8") as f:
                adapter_cfg = json.load(f)
            base_model_name = adapter_cfg.get("base_model_name_or_path") or adapter_cfg.get("base_model") or "microsoft/phi-2"

            base_model = AutoModelForCausalLM.from_pretrained(
                base_model_name,
                dtype=dtype,
                device_map=device_map,
                low_cpu_mem_usage=low_cpu_mem_usage,
                trust_remote_code=True,
            )
            
            # DEBUG: Inspect the safetensors file before loading
            adapter_file = os.path.join(model_path, "adapter_model.safetensors")
            if os.path.exists(adapter_file):
                size = os.path.getsize(adapter_file)
                print(f"DEBUG: adapter_model.safetensors size: {size} bytes")
                with open(adapter_file, "rb") as f:
                    header = f.read(100)
                    print(f"DEBUG: adapter_model.safetensors header: {header}")
            else:
                print(f"DEBUG: adapter_model.safetensors NOT FOUND at {adapter_file}")

            self.model = PeftModel.from_pretrained(base_model, model_path)
        else:
            self.model = AutoModelForCausalLM.from_pretrained(
                model_path,
                dtype=dtype,
                device_map=device_map,
                low_cpu_mem_usage=low_cpu_mem_usage,
                trust_remote_code=True,
            )

        self.model.eval()
        
    def extract_resume(self, resume_text: str) -> Dict:
        """Extract structured information from resume"""
        
        prompt = f"""Instruction: Extract structured information from the resume. Return valid JSON with fields: name, email, phone, skills, experience, education, certifications.

Input:
{resume_text}

Output:"""
        
        # Speed Optimization: Drastic reduction to 200 to GUARANTEE finish before Heroku 30s timeout
        output = self._generate(prompt, max_length=200)
        return self._parse_json_output(output)
    
    def match_resume_to_job(self, resume_text: str, job_description: str) -> Dict:
        """Match resume to job description"""
        
        prompt = f"""Instruction: Compare the resume against the job description and provide a match score (0-100) with reasoning. Return valid JSON with fields: match_score, matching_skills, missing_skills, recommendation.

Input:
Resume:
{resume_text}

Job Description:
{job_description}

Output:"""

        # Use a lower temperature to improve format adherence.
        output = self._generate(prompt, max_length=256, temperature=0.3)
        return self._parse_json_output(output)
    
    def _generate(self, prompt: str, max_length: int = 512, temperature: float = 0.7) -> str:
        """Generate text from prompt"""

        # When using device_map="auto", pick the device of the first parameter.
        input_device = next(iter(self.model.parameters())).device
        tokenized = self.tokenizer(prompt, return_tensors="pt")
        tokenized = {k: v.to(input_device) for k, v in tokenized.items()}
        input_len = tokenized["input_ids"].shape[1]

        # Interpret max_length as a generation budget (max_new_tokens) for backward compat.
        # Cap at 150 to be safe for CPU inference latency.
        max_new_tokens = max(32, min(150, int(max_length)))
        with torch.inference_mode():
            sequences = self.model.generate(
                **tokenized,
                max_new_tokens=max_new_tokens,
                min_new_tokens=8,
                temperature=temperature,
                top_p=0.95,
                num_beams=1,
                do_sample=True,
                pad_token_id=self.tokenizer.pad_token_id,
                eos_token_id=self.tokenizer.eos_token_id,
            )

        # Decode ONLY the generated continuation; avoids returning an empty string when the
        # prompt already contains the delimiter text (e.g., "Output:").
        gen_tokens = sequences[0][input_len:]
        gen_text = self.tokenizer.decode(gen_tokens, skip_special_tokens=True).strip()
        if gen_text:
            return gen_text

        # Fallback: full decode so callers can see what happened.
        full_text = self.tokenizer.decode(sequences[0], skip_special_tokens=True)
        return full_text.strip()
    
    def _parse_json_output(self, output: str) -> Dict:
        """Extract JSON from model output"""

        def _split_skills(v: Union[str, List[str], None]) -> List[str]:
            if v is None:
                return []
            if isinstance(v, list):
                return [str(s).strip() for s in v if str(s).strip()]
            v = str(v).strip()
            if not v or v.lower() in {"none", "n/a", "na"}:
                return []
            return [s.strip() for s in v.split(",") if s.strip()]

        def _normalize(d: Dict) -> Dict:
            if not isinstance(d, dict):
                return {"raw_output": output}

            # Normalize match_score to 0-100
            if "match_score" in d:
                try:
                    score = d["match_score"]
                    if isinstance(score, str):
                        score = float(re.findall(r"[0-9]*\.?[0-9]+", score)[0])
                    else:
                        score = float(score)
                    if score <= 1.0:
                        score *= 100.0
                    d["match_score"] = score
                except Exception:
                    pass

            # Normalize skills fields to lists
            if "matching_skills" in d:
                d["matching_skills"] = _split_skills(d.get("matching_skills"))
            if "missing_skills" in d:
                d["missing_skills"] = _split_skills(d.get("missing_skills"))

            # Preserve raw output for debugging
            d.setdefault("raw_output", output)
            return d
        
        try:
            # Try to find JSON in the output
            json_match = re.search(r'\{.*\}', output, re.DOTALL)
            if json_match:
                json_str = json_match.group(0)
                return _normalize(json.loads(json_str))
        except json.JSONDecodeError:
            pass

        # Fallback: parse simple key:value lines (common when the model doesn't emit JSON).
        # Example:
        # match_score: 0.85
        # matching_skills: Python, TensorFlow
        if isinstance(output, str):
            kv = {}
            for raw_line in output.splitlines():
                line = raw_line.strip()
                if not line or ":" not in line:
                    continue
                key, value = line.split(":", 1)
                key = key.strip().strip('"').strip("'").lower()
                value = value.strip().strip('"').strip("'")
                if not key:
                    continue
                kv[key] = value

            if kv:
                # Normalize known fields
                if "match_score" in kv:
                    try:
                        score = float(re.findall(r"[0-9]*\.?[0-9]+", kv["match_score"])[0])
                        if score <= 1.0:
                            score *= 100.0
                        kv["match_score"] = score
                    except Exception:
                        pass

                if "matching_skills" in kv:
                    kv["matching_skills"] = _split_skills(kv["matching_skills"])
                if "missing_skills" in kv:
                    kv["missing_skills"] = _split_skills(kv["missing_skills"])

                # Keep a copy of the original raw output for debugging
                kv["raw_output"] = output
                return kv

        # Fallback: try to parse a match score from plain text.
        m = re.search(r"match\s*score\s*[:=]\s*([0-9]*\.?[0-9]+)", output or "", flags=re.IGNORECASE)
        if m:
            score = float(m.group(1))
            if score <= 1.0:
                score *= 100.0
            return {"match_score": score, "raw_output": output}
        
        # Return structured response if parsing fails
        return {"error": "Failed to parse output", "raw_output": output}
    
    def batch_extract(self, resumes: List[str]) -> List[Dict]:
        """Extract from multiple resumes"""
        results = []
        for i, resume in enumerate(resumes):
            print(f"Processing resume {i+1}/{len(resumes)}...")
            results.append(self.extract_resume(resume))
        return results
    
    def batch_match(self, resume_pairs: List[tuple]) -> List[Dict]:
        """Match multiple resume-job pairs"""
        results = []
        for i, (resume, job) in enumerate(resume_pairs):
            print(f"Processing pair {i+1}/{len(resume_pairs)}...")
            results.append(self.match_resume_to_job(resume, job))
        return results


# Flask API for serving predictions
def create_api(model_path: str = "models/checkpoints/final"):
    """Create Flask API for inference"""
    
    from flask import Flask, request, jsonify
    from flask_cors import CORS
    
    app = Flask(__name__)
    CORS(app) # Enable CORS for all routes
    engine = ResumeInferenceEngine(model_path)
    
    @app.route("/extract", methods=["POST"])
    def extract():
        """Extract information from resume"""
        data = request.json
        resume = data.get("resume", "")
        
        if not resume:
            return jsonify({"error": "Resume text required"}), 400
        
        result = engine.extract_resume(resume)
        return jsonify(result)
    
    @app.route("/match", methods=["POST"])
    def match():
        """Match resume to job description"""
        data = request.json
        resume = data.get("resume", "")
        job = data.get("job_description", "")
        
        if not resume or not job:
            return jsonify({"error": "Resume and job description required"}), 400
        
        result = engine.match_resume_to_job(resume, job)
        return jsonify(result)
    
    @app.route("/health", methods=["GET"])
    def health():
        return jsonify({"status": "healthy"})
    
    return app


def main():
    import argparse
    
    parser = argparse.ArgumentParser()
    parser.add_argument("--mode", default="cli", help="Mode: cli, api, or batch")
    parser.add_argument("--model-path", default="models/checkpoints/final", help="Path to model")
    parser.add_argument("--task", choices=["extract", "match"], default="extract")
    parser.add_argument("--resume-file", help="Path to resume file")
    parser.add_argument("--job-file", help="Path to job description file")
    parser.add_argument("--port", type=int, default=5000, help="API port")
    
    args = parser.parse_args()
    
    engine = ResumeInferenceEngine(args.model_path)
    
    if args.mode == "cli":
        if args.task == "extract":
            with open(args.resume_file) as f:
                resume = f.read()
            result = engine.extract_resume(resume)
            print(json.dumps(result, indent=2))
        
        elif args.task == "match":
            with open(args.resume_file) as f:
                resume = f.read()
            with open(args.job_file) as f:
                job = f.read()
            result = engine.match_resume_to_job(resume, job)
            print(json.dumps(result, indent=2))
    
    elif args.mode == "api":
        app = create_api(args.model_path)
        print(f"Starting API on port {args.port}...")
        app.run(host="0.0.0.0", port=args.port, debug=False)


if __name__ == "__main__":
    main()