semantic-search / search_utils.py
Testys's picture
Update search_utils.py
b1812af
raw
history blame
24.5 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
# 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
self.api_cache = {}
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 nested directory handling"""
zip_path = Path("metadata_shards.zip")
# Check if we need to unzip by looking for parquet files in any subdirectory
if not any(self.shard_dir.rglob("*.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:
# Check for nested directory structure in zip
zip_root = self._get_zip_root(zip_ref)
# Extract while preserving structure
zip_ref.extractall(self.shard_dir)
# Handle nested directory if exists
if zip_root:
nested_dir = self.shard_dir / zip_root
if nested_dir.exists():
# Move files up from nested directory
self._flatten_directory(nested_dir, self.shard_dir)
nested_dir.rmdir()
# Verify extraction
parquet_files = list(self.shard_dir.rglob("*.parquet"))
if not parquet_files:
raise RuntimeError("Extraction completed but no parquet files found")
logger.info(f"Found {len(parquet_files)} parquet files after extraction")
except Exception as e:
logger.error(f"Failed to extract zip file: {str(e)}")
self._clean_failed_extraction()
raise
def _get_zip_root(self, zip_ref):
"""Identify common root directory in zip file"""
try:
first_file = zip_ref.namelist()[0]
if '/' in first_file:
return first_file.split('/')[0]
return ""
except Exception as e:
logger.warning(f"Error detecting zip root: {str(e)}")
return ""
def _flatten_directory(self, src_dir, dest_dir):
"""Move files from nested directory to destination"""
for item in src_dir.iterdir():
if item.is_dir():
self._flatten_directory(item, dest_dir)
item.rmdir()
else:
target = dest_dir / item.name
if target.exists():
target.unlink()
item.rename(target)
def _clean_failed_extraction(self):
"""Remove any extracted files after failed attempt"""
logger.info("Cleaning up failed extraction")
for item in self.shard_dir.iterdir():
if item.is_dir():
shutil.rmtree(item)
else:
item.unlink()
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"])
def _resolve_paper_url(self, title):
"""Find paper URL using multiple strategies"""
# Check cache first
if title in self.api_cache:
return self.api_cache[title]
links = {}
# Try arXiv first
arxiv_url = self._get_arxiv_url(title)
if arxiv_url:
links["arxiv"] = arxiv_url
# Attempt to get a direct link using Semantic Scholar's API
semantic_url = self._get_semantic_scholar_url(title)
if semantic_url:
links["semantic_search"] = semantic_url
# Fallback to Google Scholar search
scholar_url = f"https://scholar.google.com/scholar?q={quote(title)}"
links["google"] = scholar_url
self.api_cache[title] = links
return links
def _get_arxiv_url(self, title):
"""Search arXiv API for paper"""
try:
response = requests.get(
"http://export.arxiv.org/api/query",
params={
"search_query": f'ti:"{title}"',
"max_results": 1,
"sortBy": "relevance"
},
timeout=5
)
response.raise_for_status()
# Parse XML response
from xml.etree import ElementTree as ET
root = ET.fromstring(response.content)
entry = root.find('{http://www.w3.org/2005/Atom}entry')
if entry is not None:
arxiv_id = entry.find('{http://www.w3.org/2005/Atom}id').text
return arxiv_id.replace('http:', 'https:') # Force HTTPS
except Exception as e:
logger.error(f"arXiv API failed for '{title}': {str(e)}")
return None
def _get_semantic_scholar_url(self, title):
"""Search Semantic Scholar API for a paper by title and return its URL."""
try:
response = requests.get(
"https://api.semanticscholar.org/graph/v1/paper/search",
params={
"query": title,
"limit": 1,
"fields": "paperId,url,title"
},
timeout=5
)
response.raise_for_status() # This raises for 429 or other errors
data = response.json()
if "data" in data and len(data["data"]) > 0:
paper = data["data"][0]
if paper.get("url"):
return paper["url"]
elif paper.get("paperId"):
return f"https://www.semanticscholar.org/paper/{paper['paperId']}"
except requests.exceptions.HTTPError as http_err:
if response.status_code == 429:
# logger.error(f"Rate limit exceeded for Semantic Scholar API for '{title}'. Falling back.")
# Optionally, add a sleep delay here for backoff
time.sleep(1) # simple backoff delay; consider exponential backoff
else:
# logger.error(f"Semantic Scholar API failed for '{title}': {http_err}")
except Exception as e:
# logger.error(f"Semantic Scholar API failed for '{title}': {e}")
return None
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}")
results['source'] = results['title'].apply(
lambda title: self._format_source_links(
self.metadata_mgr._resolve_paper_url(title)
)
)
# 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"])
def _format_source_links(self, links):
"""Generate an HTML snippet for the available source links."""
html_parts = []
if "arxiv" in links:
html_parts.append(
f"<a class='source-link' href='{links['arxiv']}' target='_blank' rel='noopener noreferrer'> πŸ“œ arXiv</a>"
)
if "semantic" in links:
html_parts.append(
f"<a class='source-link' href='{links['semantic']}' target='_blank' rel='noopener noreferrer'> 🌐 Semantic Scholar</a>"
)
if "google" in links:
html_parts.append(
f"<a class='source-link' href='{links['google']}' target='_blank' rel='noopener noreferrer'> πŸ” Google Scholar</a>"
)
return " | ".join(html_parts)