search_utils.py

import numpy as np
import pandas as pd
import faiss
import zipfile
import logging
from pathlib import Path
from sentence_transformers import SentenceTransformer, util
import streamlit as st
import time
import os

# Configure logging
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
    handlers=[
        logging.StreamHandler()
    ]
)
logger = logging.getLogger("MetadataManager")

class MetadataManager:
    def __init__(self):
        self.cache_dir = Path("unzipped_cache")
        self.shard_dir = self.cache_dir / "metadata_shards"
        self.shard_map = {}
        self.loaded_shards = {}
        self.total_docs = 0
        
        logger.info("Initializing MetadataManager")
        self._ensure_directories()
        self._unzip_if_needed()
        self._build_shard_map()
        logger.info(f"Total documents indexed: {self.total_docs}")
        logger.info(f"Total shards found: {len(self.shard_map)}")

    def _ensure_directories(self):
        """Create necessary directories if they don't exist"""
        self.cache_dir.mkdir(parents=True, exist_ok=True)
        self.shard_dir.mkdir(parents=True, exist_ok=True)

    def _unzip_if_needed(self):
        """Handle ZIP extraction with validation and retries"""
        zip_path = Path("metadata_shards.zip")
        
        # Check if we need to unzip
        if not any(self.shard_dir.glob("*.parquet")):
            logger.info("No parquet files found, checking for zip archive")
            
            if not zip_path.exists():
                raise FileNotFoundError(f"Metadata ZIP file not found at {zip_path}")
                
            logger.info(f"Extracting {zip_path} to {self.shard_dir}")
            try:
                with zipfile.ZipFile(zip_path, 'r') as zip_ref:
                    # Validate zip contents before extraction
                    zip_files = zip_ref.namelist()
                    if not any(fname.endswith('.parquet') for fname in zip_files):
                        raise ValueError("ZIP file contains no parquet files")
                        
                    zip_ref.extractall(self.shard_dir)
                    logger.info(f"Extracted {len(zip_files)} files")
                    
                    # Verify extraction succeeded
                    if not any(self.shard_dir.glob("*.parquet")):
                        raise RuntimeError("Extraction completed but no parquet files found")
                        
            except Exception as e:
                logger.error(f"Failed to extract zip file: {str(e)}")
                # Clean up partial extraction
                if any(self.shard_dir.iterdir()):
                    for f in self.shard_dir.glob("*"):
                        f.unlink()
                raise
                
        else:
            logger.info("Parquet files already exist in cache directory")

    def _build_shard_map(self):
        """Create validated index range to shard mapping"""
        logger.info("Building shard map from parquet files")
        parquet_files = list(self.shard_dir.glob("*.parquet"))
        
        if not parquet_files:
            raise FileNotFoundError("No parquet files found after extraction")
            
        # Sort files by numerical order
        parquet_files = sorted(parquet_files, key=lambda x: int(x.stem.split("_")[1]))
        
        # Track expected next index
        expected_start = 0
        
        for f in parquet_files:
            try:
                parts = f.stem.split("_")
                if len(parts) != 3:
                    raise ValueError("Invalid filename format")
                    
                start = int(parts[1])
                end = int(parts[2])
                
                # Validate continuity
                if start != expected_start:
                    raise ValueError(f"Non-contiguous shard start: expected {expected_start}, got {start}")
                    
                # Validate range
                if end <= start:
                    raise ValueError(f"Invalid shard range: {start}-{end}")
                    
                self.shard_map[(start, end)] = f.name
                self.total_docs = end + 1
                expected_start = end + 1
                
                logger.debug(f"Mapped shard {f.name}: indices {start}-{end}")
                
            except Exception as e:
                logger.error(f"Error processing shard {f.name}: {str(e)}")
                raise RuntimeError("Invalid shard structure") from e
                
        logger.info(f"Validated {len(self.shard_map)} continuous shards")
        logger.info(f"Total document count: {self.total_docs}")
        
        # Log shard statistics
        logger.info(f"Shard map built with {len(self.shard_map)} shards")
        logger.info(f"Total document count: {self.total_docs}")
        
        # Validate shard boundaries for gaps or overlaps
        sorted_ranges = sorted(self.shard_map.keys())
        for i in range(1, len(sorted_ranges)):
            prev_end = sorted_ranges[i-1][1]
            curr_start = sorted_ranges[i][0]
            if curr_start != prev_end + 1:
                logger.warning(f"Gap or overlap detected between shards: {prev_end} to {curr_start}")

    def get_metadata(self, global_indices):
        """Retrieve metadata with validation"""
        # Check for empty numpy array properly
        if isinstance(global_indices, np.ndarray) and global_indices.size == 0:
            logger.warning("Empty indices array passed to get_metadata")
            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
        
        # Convert numpy array to list for processing
        indices_list = global_indices.tolist() if isinstance(global_indices, np.ndarray) else global_indices
        logger.info(f"Retrieving metadata for {len(indices_list)} indices")
        
        # Filter valid indices
        valid_indices = [idx for idx in indices_list if 0 <= idx < self.total_docs]
        invalid_count = len(indices_list) - len(valid_indices)
        if invalid_count > 0:
            logger.warning(f"Filtered out {invalid_count} invalid indices")
        
        if not valid_indices:
            logger.warning("No valid indices remain after filtering")
            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
    
        # Group indices by shard with boundary check
        shard_groups = {}
        unassigned_indices = []
        
        for idx in valid_indices:
            found = False
            for (start, end), shard in self.shard_map.items():
                if start <= idx <= end:
                    if shard not in shard_groups:
                        shard_groups[shard] = []
                    shard_groups[shard].append(idx - start)
                    found = True
                    break
            if not found:
                unassigned_indices.append(idx)
                logger.warning(f"Index {idx} not found in any shard range")
        
        if unassigned_indices:
            logger.warning(f"Could not assign {len(unassigned_indices)} indices to any shard")
    
        # Load and process shards
        results = []
        for shard, local_indices in shard_groups.items():
            try:
                logger.info(f"Processing shard {shard} with {len(local_indices)} indices")
                start_time = time.time()
                
                if shard not in self.loaded_shards:
                    logger.info(f"Loading shard file: {shard}")
                    shard_path = self.shard_dir / shard
                    
                    # Verify file exists
                    if not shard_path.exists():
                        logger.error(f"Shard file not found: {shard_path}")
                        continue
                        
                    # Log file size
                    file_size_mb = os.path.getsize(shard_path) / (1024 * 1024)
                    logger.info(f"Shard file size: {file_size_mb:.2f} MB")
                    
                    # Attempt to read the parquet file
                    try:
                        self.loaded_shards[shard] = pd.read_parquet(
                            shard_path,
                            columns=["title", "summary", "source"]
                        )
                        logger.info(f"Successfully loaded shard {shard} with {len(self.loaded_shards[shard])} rows")
                    except Exception as e:
                        logger.error(f"Failed to read parquet file {shard}: {str(e)}")
                        
                        # Try to read file schema for debugging
                        try:
                            schema = pd.read_parquet(shard_path, engine='pyarrow').dtypes
                            logger.info(f"Parquet schema: {schema}")
                        except:
                            pass
                        continue
                
                if local_indices:
                    # Validate indices are within dataframe bounds
                    df_len = len(self.loaded_shards[shard])
                    valid_local_indices = [idx for idx in local_indices if 0 <= idx < df_len]
                    
                    if len(valid_local_indices) != len(local_indices):
                        logger.warning(f"Filtered {len(local_indices) - len(valid_local_indices)} out-of-bounds indices")
                    
                    if valid_local_indices:
                        logger.debug(f"Retrieving rows at indices: {valid_local_indices}")
                        chunk = self.loaded_shards[shard].iloc[valid_local_indices]
                        results.append(chunk)
                        logger.info(f"Retrieved {len(chunk)} records from shard {shard}")
                
                logger.info(f"Shard processing completed in {time.time() - start_time:.2f} seconds")
                    
            except Exception as e:
                logger.error(f"Error processing shard {shard}: {str(e)}", exc_info=True)
                continue
    
        # Combine results
        if results:
            combined = pd.concat(results).reset_index(drop=True)
            logger.info(f"Combined metadata: {len(combined)} records from {len(results)} shards")
            return combined
        else:
            logger.warning("No metadata records retrieved")
            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])

class SemanticSearch:
    def __init__(self):
        self.shard_dir = Path("compressed_shards")
        self.model = None
        self.index_shards = []
        self.metadata_mgr = MetadataManager()
        self.shard_sizes = []
        
        # Configure search logger
        self.logger = logging.getLogger("SemanticSearch")
        self.logger.info("Initializing SemanticSearch")
        
    @st.cache_resource
    def load_model(_self):
        return SentenceTransformer('all-MiniLM-L6-v2')

    def initialize_system(self):
        self.logger.info("Loading sentence transformer model")
        start_time = time.time()
        self.model = self.load_model()
        self.logger.info(f"Model loaded in {time.time() - start_time:.2f} seconds")
        
        self.logger.info("Loading FAISS indices")
        self._load_faiss_shards()

    def _load_faiss_shards(self):
        """Load all FAISS index shards"""
        self.logger.info(f"Searching for index files in {self.shard_dir}")
        
        if not self.shard_dir.exists():
            self.logger.error(f"Shard directory not found: {self.shard_dir}")
            return
            
        index_files = list(self.shard_dir.glob("*.index"))
        self.logger.info(f"Found {len(index_files)} index files")
        
        self.shard_sizes = []
        self.index_shards = []
        
        for shard_path in sorted(index_files):
            try:
                self.logger.info(f"Loading index: {shard_path}")
                start_time = time.time()
                
                # Log file size
                file_size_mb = os.path.getsize(shard_path) / (1024 * 1024)
                self.logger.info(f"Index file size: {file_size_mb:.2f} MB")
                
                index = faiss.read_index(str(shard_path))
                self.index_shards.append(index)
                self.shard_sizes.append(index.ntotal)
                
                self.logger.info(f"Loaded index with {index.ntotal} vectors in {time.time() - start_time:.2f} seconds")
            except Exception as e:
                self.logger.error(f"Failed to load index {shard_path}: {str(e)}")
        
        total_vectors = sum(self.shard_sizes)
        self.logger.info(f"Total loaded vectors: {total_vectors} across {len(self.index_shards)} shards")

    def _global_index(self, shard_idx, local_idx):
        """Convert local index to global index"""
        return sum(self.shard_sizes[:shard_idx]) + local_idx

    def search(self, query, top_k=5):
        """Search with validation"""
        self.logger.info(f"Searching for query: '{query}' (top_k={top_k})")
        start_time = time.time()
        
        if not query:
            self.logger.warning("Empty query provided")
            return pd.DataFrame()
            
        if not self.index_shards:
            self.logger.error("No index shards loaded")
            return pd.DataFrame()
        
        try:
            self.logger.info("Encoding query")
            query_embedding = self.model.encode([query], convert_to_numpy=True)
            self.logger.debug(f"Query encoded to shape {query_embedding.shape}")
        except Exception as e:
            self.logger.error(f"Query encoding failed: {str(e)}")
            return pd.DataFrame()
    
        all_distances = []
        all_global_indices = []
    
        # Search with index validation
        self.logger.info(f"Searching across {len(self.index_shards)} shards")
        for shard_idx, index in enumerate(self.index_shards):
            if index.ntotal == 0:
                self.logger.warning(f"Skipping empty shard {shard_idx}")
                continue
                
            try:
                shard_start = time.time()
                distances, indices = index.search(query_embedding, top_k)
                
                valid_mask = (indices[0] >= 0) & (indices[0] < index.ntotal)
                valid_indices = indices[0][valid_mask].tolist()
                valid_distances = distances[0][valid_mask].tolist()
                
                if len(valid_indices) != top_k:
                    self.logger.debug(f"Shard {shard_idx}: Found {len(valid_indices)} valid results out of {top_k}")
                
                global_indices = [self._global_index(shard_idx, idx) for idx in valid_indices]
                
                all_distances.extend(valid_distances)
                all_global_indices.extend(global_indices)
                
                self.logger.debug(f"Shard {shard_idx} search completed in {time.time() - shard_start:.3f}s")
            except Exception as e:
                self.logger.error(f"Search failed in shard {shard_idx}: {str(e)}")
                continue
    
        self.logger.info(f"Search found {len(all_global_indices)} results across all shards")
        
        # Process results
        results = self._process_results(
            np.array(all_distances), 
            np.array(all_global_indices), 
            top_k
        )
        
        self.logger.info(f"Search completed in {time.time() - start_time:.2f} seconds with {len(results)} final results")
        return results

    def _process_results(self, distances, global_indices, top_k):
        """Process raw search results into formatted DataFrame"""
        process_start = time.time()
        
        # Proper numpy array emptiness checks
        if global_indices.size == 0 or distances.size == 0:
            self.logger.warning("No search results to process")
            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
        
        try:
            # Get metadata for matched indices
            self.logger.info(f"Retrieving metadata for {len(global_indices)} indices")
            metadata_start = time.time()
            results = self.metadata_mgr.get_metadata(global_indices)
            self.logger.info(f"Metadata retrieved in {time.time() - metadata_start:.2f}s, got {len(results)} records")
            
            # Empty results check
            if len(results) == 0:
                self.logger.warning("No metadata found for indices")
                return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
                
            # Ensure distances match results length
            if len(results) != len(distances):
                self.logger.warning(f"Mismatch between distances ({len(distances)}) and results ({len(results)})")
                
                if len(results) < len(distances):
                    self.logger.info("Truncating distances array to match results length")
                    distances = distances[:len(results)]
                else:
                    # Should not happen but handle it anyway
                    self.logger.error("More results than distances - this shouldn't happen")
                    distances = np.pad(distances, (0, len(results) - len(distances)), 'constant', constant_values=1.0)
                
            # Calculate similarity scores
            self.logger.debug("Calculating similarity scores")
            results['similarity'] = 1 - (distances / 2)
            
            # Log similarity statistics
            if not results.empty:
                self.logger.debug(f"Similarity stats: min={results['similarity'].min():.3f}, " +
                                 f"max={results['similarity'].max():.3f}, " +
                                 f"mean={results['similarity'].mean():.3f}")
            
            # Deduplicate and sort results
            pre_dedup = len(results)
            results = results.drop_duplicates(subset=["title", "source"]).sort_values("similarity", ascending=False).head(top_k)
            post_dedup = len(results)
            
            if pre_dedup > post_dedup:
                self.logger.info(f"Removed {pre_dedup - post_dedup} duplicate results")
            
            self.logger.info(f"Results processed in {time.time() - process_start:.2f}s, returning {len(results)} items")
            return results.reset_index(drop=True)
            
        except Exception as e:
            self.logger.error(f"Result processing failed: {str(e)}", exc_info=True)
            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])