app/services/reranker_service.py

import logging
import torch
from typing import List, Dict
from transformers import PreTrainedTokenizer
import math
import os
import requests
# --- Import your custom model architecture ---
from app.models.expert_judge_model import ExpertJudgeCrossEncoder, get_tokenizer

from app.core import state
from app.core.config import settings

logger = logging.getLogger(__name__)

def load_reranker_model():
    """
    Loads the custom-trained ExpertJudgeCrossEncoder model. If running on a
    new server, it first downloads the model from S3.
    """
    if state.reranker_model_loaded:
        logger.info("Re-ranker model already loaded in state.")
        return True
 
    # --- ADDED: Download from S3 if file doesn't exist ---
    model_path = settings.RERANKER_MODEL_PATH
    if not os.path.exists(model_path) and settings.S3_RERANKER_URL:
        logger.info(f"Re-ranker model not found at {model_path}. Downloading from S3...")
        try:
            # Create the 'data' directory if it doesn't exist
            os.makedirs(os.path.dirname(model_path), exist_ok=True)
            with requests.get(settings.S3_RERANKER_URL, stream=True) as r:
                r.raise_for_status()
                with open(model_path, 'wb') as f:
                    for chunk in r.iter_content(chunk_size=8192):
                        f.write(chunk)
            logger.info("Successfully downloaded re-ranker model from S3.")
        except Exception as e:
            logger.exception(f"FATAL: Failed to download re-ranker model from S3: {e}")
            return False
    # --- END OF ADDITION ---
 
    base_model_name = settings.RERANKER_MODEL_NAME
    logger.info(f"Loading custom ExpertJudgeCrossEncoder from: {model_path}")
    logger.info(f"Using base model architecture: {base_model_name}")
 
    try:
        # 1. Instantiate the model structure
        model = ExpertJudgeCrossEncoder(model_name=base_model_name)
        # 2. Load the saved weights (the state_dict) into the model structure
        model.load_state_dict(torch.load(model_path, map_location=state.device))
        # 3. Set up the model for inference
        model.to(state.device)
        model.eval()
        # 4. Load the corresponding tokenizer
        tokenizer = get_tokenizer(model_name=base_model_name)
 
        # 5. Store both in the state
        state.reranker_model = model
        state.reranker_tokenizer = tokenizer
        state.reranker_model_loaded = True
        logger.info("Custom ExpertJudgeCrossEncoder model and tokenizer loaded successfully.")
        return True
    except Exception as e:
        logger.exception(f"Failed to load custom Cross-Encoder model: {e}")
        return False

def rerank_chunks(query: str, chunks: List[Dict], metadata_map: Dict) -> List[Dict]:
    """
    Re-ranks chunks using the custom ExpertJudgeCrossEncoder with the concise
    v4 input format.
    """
    if not state.reranker_model_loaded or not chunks:
        logger.warning("Re-ranker not loaded or no chunks provided. Returning original order.")
        return chunks

    logger.info(f"Re-ranking {len(chunks)} chunks with custom Expert Judge (v4 format)...")

    model = state.reranker_model
    tokenizer: PreTrainedTokenizer = state.reranker_tokenizer
    scores = []
 # --- EDIT: Create pairs of [query, passage] for the tokenizer ---
    # The BGE-Reranker model was trained on this simple format.
    query_passage_pairs = [[query, chunk.get('text', '')] for chunk in chunks]

    with torch.no_grad():
        # --- EDIT: Tokenize the pairs directly ---
        # The tokenizer will handle adding [CLS], [SEP] tokens correctly for pairs.
        encoded = tokenizer(
            query_passage_pairs,
            padding=True,
            truncation=True,
            return_tensors='pt',
            max_length=512
        )
        
        encoded = {k: v.to(state.device) for k, v in encoded.items()}
        
        logits = model(input_ids=encoded['input_ids'], attention_mask=encoded['attention_mask'])
        # scores = torch.sigmoid(logits).squeeze().cpu().numpy()
        scores = logits.squeeze().cpu().numpy()

    # Add the new, more accurate score to each chunk dictionary
    if len(chunks) == 1:
        chunks[0]['rerank_score'] = float(scores)
    else:
        for i, chunk in enumerate(chunks):
            chunk['rerank_score'] = scores[i]

    # Sort the chunks by the new score in descending order
    sorted_chunks = sorted(chunks, key=lambda x: x.get('rerank_score', 0.0), reverse=True)

    logger.info("--- Top 5 Custom Re-ranked Chunks (v4) ---")
    for i, chunk in enumerate(sorted_chunks[:100]):
        logger.info(f"   {i+1}. Chunk ID: {chunk.get('id', 'N/A')} | New Score: {chunk.get('rerank_score', 0.0):.4f}")
    logger.info("--------------------------------")

    return sorted_chunks