[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.
Python aifaissgpt-osslangchainllama-indexllmlocalstorageoffline-firstollamaprivacypythonragretrieval-augmented-generationvector-databasevector-searchvectors
feat: compare faiss
Andy Lee711fb4a77537
diff --git a/examples/compare_faiss_vs_leann.py b/examples/compare_faiss_vs_leann.pynew file mode 100644index 0000000..70487f3--- /dev/null+++ b/examples/compare_faiss_vs_leann.py@@ -0,0 +1,352 @@+#!/usr/bin/env python3+"""+Memory comparison between Faiss HNSW and LEANN HNSW backend+"""++import logging+import sys+import time+import psutil+import gc++# Setup logging+logging.basicConfig(stream=sys.stdout, level=logging.INFO)+logger = logging.getLogger(__name__)+++def get_memory_usage():+ """Get current memory usage in MB"""+ process = psutil.Process()+ return process.memory_info().rss / 1024 / 1024+++def print_memory_stats(stage: str, start_mem: float):+ """Print memory statistics"""+ current_mem = get_memory_usage()+ diff = current_mem - start_mem+ print(f"[{stage}] Memory: {current_mem:.1f} MB (+{diff:.1f} MB)")+ return current_mem+++class MemoryTracker:+ def __init__(self, name: str):+ self.name = name+ self.start_mem = get_memory_usage()+ self.stages = []++ def checkpoint(self, stage: str):+ current_mem = print_memory_stats(f"{self.name} - {stage}", self.start_mem)+ self.stages.append((stage, current_mem))+ return current_mem++ def summary(self):+ print(f"\n=== {self.name} Memory Summary ===")+ for stage, mem in self.stages:+ print(f"{stage}: {mem:.1f} MB")+ peak_mem = max(mem for _, mem in self.stages)+ print(f"Peak Memory: {peak_mem:.1f} MB")+ print(f"Total Memory Increase: {peak_mem - self.start_mem:.1f} MB")+ return peak_mem+++def test_faiss_hnsw():+ """Test Faiss HNSW Vector Store"""+ print("\n" + "=" * 50)+ print("TESTING FAISS HNSW VECTOR STORE")+ print("=" * 50)++ try:+ import faiss+ from llama_index.core import (+ SimpleDirectoryReader,+ VectorStoreIndex,+ StorageContext,+ Settings,+ )+ from llama_index.vector_stores.faiss import FaissVectorStore+ from llama_index.embeddings.huggingface import HuggingFaceEmbedding+ except ImportError as e:+ print(f"❌ Missing dependencies for Faiss test: {e}")+ print("Please install:")+ print(" pip install faiss-cpu")+ print(" pip install llama-index-vector-stores-faiss")+ print(" pip install llama-index-embeddings-huggingface")+ return {+ "build_time": float("inf"),+ "peak_memory": float("inf"),+ "error": str(e),+ }++ tracker = MemoryTracker("Faiss HNSW")++ # Import and setup+ tracker.checkpoint("Initial")++ tracker.checkpoint("After imports")++ # Setup embedding model (same as LEANN)+ embed_model = HuggingFaceEmbedding(model_name="facebook/contriever")+ Settings.embed_model = embed_model++ tracker.checkpoint("After embedding model setup")++ # Create Faiss index+ d = 768 # facebook/contriever embedding dimension+ faiss_index = faiss.IndexHNSWFlat(d, 32) # M=32 same as LEANN+ faiss_index.hnsw.efConstruction = 64 # same as LEANN complexity++ tracker.checkpoint("After Faiss index creation")++ # Load documents+ documents = SimpleDirectoryReader(+ "examples/data",+ recursive=True,+ encoding="utf-8",+ required_exts=[".pdf", ".txt", ".md"],+ ).load_data()++ tracker.checkpoint("After document loading")++ # Create vector store and index+ vector_store = FaissVectorStore(faiss_index=faiss_index)+ storage_context = StorageContext.from_defaults(vector_store=vector_store)++ # Build index+ print("Building Faiss HNSW index...")+ start_time = time.time()+ index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)+ build_time = time.time() - start_time++ tracker.checkpoint("After index building")++ # Save index+ index.storage_context.persist("./storage_faiss")+ tracker.checkpoint("After index saving")++ # Test queries+ query_engine = index.as_query_engine(similarity_top_k=20)++ print("Running queries...")+ queries = [+ "什么是盘古大模型以及盘古开发过程中遇到了什么阴暗面,任务令一般在什么城市颁发",+ "What is LEANN and how does it work?",+ "华为诺亚方舟实验室的主要研究内容",+ ]++ for i, query in enumerate(queries):+ start_time = time.time()+ response = query_engine.query(query)+ query_time = time.time() - start_time+ print(f"Query {i + 1} time: {query_time:.3f}s")+ tracker.checkpoint(f"After query {i + 1}")++ peak_memory = tracker.summary()++ # Clean up+ del index, vector_store, storage_context, faiss_index+ gc.collect()++ return {"build_time": build_time, "peak_memory": peak_memory, "tracker": tracker}+++def test_leann_hnsw():+ """Test LEANN HNSW Backend"""+ print("\n" + "=" * 50)+ print("TESTING LEANN HNSW BACKEND")+ print("=" * 50)++ tracker = MemoryTracker("LEANN HNSW")++ # Import and setup+ tracker.checkpoint("Initial")++ from llama_index.core import SimpleDirectoryReader+ from llama_index.core.node_parser import SentenceSplitter+ from leann.api import LeannBuilder, LeannChat+ from pathlib import Path++ tracker.checkpoint("After imports")++ # Load and parse documents+ documents = SimpleDirectoryReader(+ "examples/data",+ recursive=True,+ encoding="utf-8",+ required_exts=[".pdf", ".txt", ".md"],+ ).load_data()++ tracker.checkpoint("After document loading")++ # Parse into chunks+ node_parser = SentenceSplitter(+ chunk_size=256, chunk_overlap=20, separator=" ", paragraph_separator="\n\n"+ )++ all_texts = []+ for doc in documents:+ nodes = node_parser.get_nodes_from_documents([doc])+ for node in nodes:+ all_texts.append(node.get_content())++ tracker.checkpoint("After text chunking")++ # Build LEANN index+ INDEX_DIR = Path("./test_leann_comparison")+ INDEX_PATH = str(INDEX_DIR / "comparison.leann")++ # Clean up previous index+ import shutil++ if INDEX_DIR.exists():+ shutil.rmtree(INDEX_DIR)++ builder = LeannBuilder(+ backend_name="hnsw",+ embedding_model="facebook/contriever",+ graph_degree=32,+ complexity=64,+ is_compact=True,+ is_recompute=True,+ num_threads=1,+ )++ tracker.checkpoint("After builder setup")++ print("Building LEANN HNSW index...")+ start_time = time.time()++ for chunk_text in all_texts:+ builder.add_text(chunk_text)++ builder.build_index(INDEX_PATH)+ build_time = time.time() - start_time++ tracker.checkpoint("After index building")++ # Ensure any existing embedding server is updated with new data+ print("Ensuring embedding server data synchronization...")+ meta_file = INDEX_DIR / "comparison.leann.meta.json"+ if meta_file.exists():+ from leann.embedding_server_manager import (+ _check_port,+ _update_server_meta_path,+ _check_server_meta_path,+ )++ port = 5557 # Default port for HNSW backend+ if _check_port(port):+ print(f"Updating server meta path to: {meta_file}")+ success = _update_server_meta_path(port, str(meta_file))+ if success:+ print("✅ Server meta path updated successfully")+ else:+ print("❌ Failed to update server meta path")++ # Verify the update and ensure server is ready+ time.sleep(2) # Give server time to reload data+ max_retries = 5+ for retry in range(max_retries):+ if _check_server_meta_path(port, str(meta_file)):+ print("✅ Server meta path verification successful")+ break+ else:+ print(+ f"⏳ Server meta path verification failed (attempt {retry + 1}/{max_retries})"+ )+ if retry < max_retries - 1:+ time.sleep(1)+ else:+ print(+ "❌ Server meta path verification failed after all retries - may cause query issues"+ )+ else:+ print("No existing server found on port 5557")++ # Test queries+ chat = LeannChat(+ index_path=INDEX_PATH, llm_config={"type": "simulated", "model": "test"}+ )++ tracker.checkpoint("After chat setup")++ print("Running queries...")+ queries = [+ "什么是盘古大模型以及盘古开发过程中遇到了什么阴暗面,任务令一般在什么城市颁发",+ "What is LEANN and how does it work?",+ "华为诺亚方舟实验室的主要研究内容",+ ]++ for i, query in enumerate(queries):+ start_time = time.time()+ response = chat.ask(+ query, top_k=20, recompute_beighbor_embeddings=True, complexity=32+ )+ query_time = time.time() - start_time+ print(f"Query {i + 1} time: {query_time:.3f}s")+ tracker.checkpoint(f"After query {i + 1}")++ peak_memory = tracker.summary()++ # Clean up+ del chat, builder+ if INDEX_DIR.exists():+ shutil.rmtree(INDEX_DIR)+ gc.collect()++ return {"build_time": build_time, "peak_memory": peak_memory, "tracker": tracker}+++def main():+ """Run comparison tests"""+ print("Memory Usage Comparison: Faiss HNSW vs LEANN HNSW")+ print("=" * 60)++ # Test Faiss HNSW+ faiss_results = test_faiss_hnsw()++ # Force garbage collection+ gc.collect()+ time.sleep(2)++ # Test LEANN HNSW+ leann_results = test_leann_hnsw()++ # Final comparison+ print("\n" + "=" * 60)+ print("FINAL COMPARISON")+ print("=" * 60)++ print(f"Faiss HNSW:")+ if "error" in faiss_results:+ print(f" ❌ Failed: {faiss_results['error']}")+ else:+ print(f" Build Time: {faiss_results['build_time']:.3f}s")+ print(f" Peak Memory: {faiss_results['peak_memory']:.1f} MB")++ print(f"\nLEANN HNSW:")+ print(f" Build Time: {leann_results['build_time']:.3f}s")+ print(f" Peak Memory: {leann_results['peak_memory']:.1f} MB")++ # Calculate improvements only if Faiss test succeeded+ if "error" not in faiss_results:+ time_ratio = faiss_results["build_time"] / leann_results["build_time"]+ memory_ratio = faiss_results["peak_memory"] / leann_results["peak_memory"]++ print(f"\nLEANN vs Faiss:")+ print(+ f" Build Time: {time_ratio:.2f}x {'faster' if time_ratio > 1 else 'slower'}"+ )+ print(+ f" Memory Usage: {memory_ratio:.2f}x {'less' if memory_ratio > 1 else 'more'}"+ )++ print(+ f"\nMemory Savings: {faiss_results['peak_memory'] - leann_results['peak_memory']:.1f} MB"+ )+ else:+ print(f"\n✅ LEANN HNSW ran successfully!")+ print(f"📊 LEANN Memory Usage: {leann_results['peak_memory']:.1f} MB")+++if __name__ == "__main__":+ main()