semantic-search / search_utils.py
Testys's picture
Update search_utils.py
4709458 verified
raw
history blame
14.9 kB
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
from urllib.parse import quote
import requests
import shutil
import concurrent.futures
from functools import lru_cache
# 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.metadata_path = Path("combined.parquet")
self.df = None
self.total_docs = 0
logger.info("Initializing MetadataManager")
self._load_metadata()
logger.info(f"Total documents indexed: {self.total_docs}")
def _load_metadata(self):
"""Load the combined parquet file directly"""
logger.info("Loading metadata from combined.parquet")
try:
# Load the parquet file
self.df = pd.read_parquet(self.metadata_path)
# Clean and format the data
self.df['source'] = self.df['source'].apply(
lambda x: [
url.strip()
for url in str(x).split(';')
if url.strip()
]
)
self.total_docs = len(self.df)
logger.info(f"Successfully loaded {self.total_docs} documents")
except Exception as e:
logger.error(f"Failed to load metadata: {str(e)}")
raise
def get_metadata(self, global_indices):
"""Retrieve metadata for given indices with deduplication by title"""
if isinstance(global_indices, np.ndarray) and global_indices.size == 0:
return pd.DataFrame(columns=["title", "summary", 'authors', "similarity", "source"])
try:
# Directly index the DataFrame
results = self.df.iloc[global_indices].copy()
# Deduplicate by title to avoid near-duplicate results
if len(results) > 1:
results = results.drop_duplicates(subset=["title"])
return results
except Exception as e:
logger.error(f"Metadata retrieval failed: {str(e)}")
return pd.DataFrame(columns=["title", "summary", "similarity", "source", 'authors'])
class SemanticSearch:
def __init__(self):
self.shard_dir = Path("compressed_shards")
self.model = None
self.index_shards = []
self.metadata_mgr = MetadataManager()
self.shard_sizes = []
self.cumulative_offsets = None
self.total_vectors = 0
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 FAISS shards concurrently and precompute cumulative offsets for global indexing."""
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 = sorted(self.shard_dir.glob("*.index"))
self.logger.info(f"Found {len(index_files)} index files")
self.index_shards = []
self.shard_sizes = []
with concurrent.futures.ThreadPoolExecutor() as executor:
future_to_file = {
executor.submit(self._load_single_index, shard_path): shard_path
for shard_path in index_files
}
for future in concurrent.futures.as_completed(future_to_file):
shard_path = future_to_file[future]
try:
index, size = future.result()
if index is not None:
self.index_shards.append(index)
self.shard_sizes.append(size)
self.logger.info(f"Loaded index {shard_path.name} with {size} vectors")
except Exception as e:
self.logger.error(f"Error loading index {shard_path}: {str(e)}")
self.total_vectors = sum(self.shard_sizes)
self.logger.info(f"Total loaded vectors: {self.total_vectors} across {len(self.index_shards)} shards")
self.cumulative_offsets = np.cumsum([0] + self.shard_sizes)
def _load_single_index(self, shard_path):
"""Load a single FAISS index shard."""
self.logger.info(f"Loading index: {shard_path}")
start_time = time.time()
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))
size = index.ntotal
self.logger.info(f"Index loaded in {time.time() - start_time:.2f} seconds")
return index, size
def _global_index(self, shard_idx, local_idx):
"""Convert a local index (within a shard) to a global index using precomputed offsets with validation."""
if shard_idx < 0 or shard_idx >= len(self.index_shards):
self.logger.error(f"Invalid shard index: {shard_idx}")
return -1
if local_idx < 0 or local_idx >= self.shard_sizes[shard_idx]:
self.logger.warning(f"Local index {local_idx} may be out of bounds for shard {shard_idx}")
return int(self.cumulative_offsets[shard_idx] + local_idx)
def search(self, query, top_k=5):
"""Search for a query using parallel FAISS shard search with normalized vectors for proper cosine similarity."""
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)
# Normalize query embedding for proper cosine similarity comparison
query_embedding = query_embedding / np.linalg.norm(query_embedding, axis=1, keepdims=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 = []
with concurrent.futures.ThreadPoolExecutor() as executor:
futures = {
executor.submit(self._search_shard, shard_idx, index, query_embedding, top_k): shard_idx
for shard_idx, index in enumerate(self.index_shards)
}
for future in concurrent.futures.as_completed(futures):
result = future.result()
if result is not None:
distances_part, global_indices_part = result
all_distances.extend(distances_part)
all_global_indices.extend(global_indices_part)
# If no results found across all shards
if not all_global_indices:
self.logger.warning("No results found across any shards")
return pd.DataFrame(columns=["title", "summary", "source", "authors", "similarity"])
self.logger.info(f"Search found {len(all_global_indices)} results across all shards")
# Sort all results by distance before processing
combined = list(zip(all_distances, all_global_indices))
combined.sort(reverse=True) # Sort by distance (higher is better for cosine similarity)
sorted_distances, sorted_indices = zip(*combined)
# Limit to top-k across all shards
top_distances = np.array(sorted_distances[:top_k])
top_indices = np.array(sorted_indices[:top_k])
results = self._process_results(top_distances, top_indices, top_k)
self.logger.info(f"Search completed in {time.time() - start_time:.2f} seconds with {len(results)} final results")
return results
def _search_shard(self, shard_idx, index, query_embedding, top_k):
"""Search a single FAISS shard for the query embedding with proper error handling."""
if index.ntotal == 0:
self.logger.warning(f"Skipping empty shard {shard_idx}")
return None
try:
shard_start = time.time()
distances, indices = index.search(query_embedding, top_k)
# Filter out invalid indices (-1 is returned by FAISS for insufficient results)
valid_mask = (indices[0] >= 0) & (indices[0] < index.ntotal)
valid_indices = indices[0][valid_mask]
valid_distances = distances[0][valid_mask]
if len(valid_indices) == 0:
self.logger.debug(f"Shard {shard_idx}: No valid results found")
return None
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]
# Filter out any invalid global indices (could happen if _global_index validation fails)
valid_global = [(d, i) for d, i in zip(valid_distances, global_indices) if i >= 0]
if not valid_global:
return None
final_distances, final_indices = zip(*valid_global)
self.logger.debug(f"Shard {shard_idx} search completed in {time.time() - shard_start:.3f}s")
return final_distances, final_indices
except Exception as e:
self.logger.error(f"Search failed in shard {shard_idx}: {str(e)}")
return None
def _process_results(self, distances, global_indices, top_k):
"""Process raw search results with correct similarity calculation for cosine similarity."""
process_start = time.time()
if global_indices.size == 0 or distances.size == 0:
self.logger.warning("No search results to process")
return pd.DataFrame(columns=["title", "summary", "source", "authors", "similarity"])
try:
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")
if len(results) == 0:
self.logger.warning("No metadata found for indices")
return pd.DataFrame(columns=["title", "summary", "source", "authors", "similarity"])
# Handle distance-results alignment
if len(results) != len(distances):
self.logger.warning(f"Mismatch between distances ({len(distances)}) and results ({len(results)})")
min_len = min(len(results), len(distances))
results = results.iloc[:min_len]
distances = distances[:min_len]
# For inner product with normalized vectors, similarity is directly the distance
# (FAISS IP search already returns higher scores for more similar items)
results['similarity'] = 1 - (distances/2)
# Deduplicate and sort
required_columns = ["title", "summary", "authors", "source", "similarity"]
pre_dedup = len(results)
results = (
results.drop_duplicates(subset=["title", "authors"])
.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")
return results[required_columns].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", "authors", "similarity"])
def search_with_threshold(self, query, top_k=10, similarity_threshold=0.6):
"""
Search with a fixed similarity threshold, returning only results above the threshold.
For cosine similarity with normalized vectors, threshold should be between 0 and 1.
"""
# Get more results initially to ensure we have enough after filtering
initial_results = self.search(query, top_k=top_k*2)
if initial_results.empty:
return initial_results
# Filter by similarity threshold
filtered_results = initial_results[initial_results['similarity'] >= similarity_threshold]
# Return top-k of filtered results
return filtered_results.head(top_k).reset_index(drop=True)
def search_with_adaptive_threshold(self, query, top_k=10, percentile=75):
"""
Search with an adaptive threshold based on the distribution of similarity scores.
Returns results above the specified percentile of similarity scores.
"""
# Get more results initially to determine distribution
initial_results = self.search(query, top_k=top_k*3)
if initial_results.empty or len(initial_results) < 2:
return initial_results
# Calculate threshold based on percentile of similarity scores
threshold = np.percentile(initial_results['similarity'], percentile)
self.logger.info(f"Adaptive threshold set to {threshold:.4f} (percentile: {percentile})")
# Filter results above threshold
filtered_results = initial_results[initial_results['similarity'] >= threshold]
# Return top-k of filtered results
return filtered_results.head(top_k).reset_index(drop=True)