[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
upd readme stats
yichuan520030910320071945877578
diff --git a/README.md b/README.mdindex 70a923e..b6e6d2b 100755--- a/README.md+++ b/README.md@@ -70,10 +70,11 @@ uv syncfrom leann.api import LeannBuilder, LeannSearcher# 1. Build index (no embeddings stored!)-builder = LeannBuilder(backend_name="diskann")+builder = LeannBuilder(backend_name="hnsw")builder.add_text("Python is a powerful programming language")builder.add_text("Machine learning transforms industries")builder.add_text("Neural networks process complex data")+builder.add_text("Leann is a great storage saving engine for RAG on your macbook")builder.build_index("knowledge.leann")# 2. Search with real-time embeddings@@ -320,24 +321,18 @@ python examples/compare_faiss_vs_leann.py| Metric | Faiss HNSW | LEANN HNSW | **Improvement** ||--------|------------|-------------|-----------------|-| **Peak Memory** | 887.0 MB | 618.2 MB | **1.4x less** (268.8 MB saved) || **Storage Size** | 5.5 MB | 0.5 MB | **11.4x smaller** (5.0 MB saved) |#### 📈 Key Takeaways-- **🧠 Memory Efficiency**: LEANN uses **30% less memory** during index building and querying+- **💾 Storage Optimization**: LEANN requires **91% less storage** for the same dataset-- **🔄 On-demand Computing**: Storage savings come from computing embeddings at query time instead of pre-storing them-- **⚖️ Fair Comparison**: Both systems tested on identical hardware with the same 2,573 document dataset++- **⚖️ Fair Comparison**: Both systems tested on identical hardware with the same 2,573 document dataset and the same embedding model and chunk method> **Note**: Results may vary based on dataset size, hardware configuration, and query patterns. The comparison excludes text storage to focus purely on index structures.-### Run the comparison--```bash-python examples/compare_faiss_vs_leann.py-```*Benchmark results obtained on Apple Silicon with consistent environmental conditions*@@ -376,13 +371,13 @@ The script will print the recall and search time for each query, followed by the### Storage Usage Comparison-| System | DPR(2.1M docs) | RPJ-wiki(60M docs) | Chat history() |+| System | DPR(2.1M docs) | RPJ-wiki(60M docs) | Chat history(5K messages) || --------------------- | ---------------- | ---------------- | ---------------- |-| Traditional Vector DB | 3.8 GB | 201 GB | TBD |-| **LEANN** | **324 MB** | **6 GB** | **TBD** |-| **Reduction** | **91% smaller** | **97% smaller** | **TBD** |+| Traditional Vector DB | 3.8 GB | 201 GB | **22.8 MB** |+| **LEANN** | **324 MB** | **6 GB** | **0.78 MB** |+| **Reduction** | **91% smaller** | **97% smaller** | **97% smaller** |-### Memory Usage Comparison+<!-- ### Memory Usage Comparison| System j | DPR(2M docs) | RPJ-wiki(60M docs) | Chat history() || --------------------- | ---------------- | ---------------- | ---------------- |@@ -395,7 +390,7 @@ The script will print the recall and search time for each query, followed by the| Backend | Index Size | Query Time | Recall@3 || ------------------- | ---------- | ---------- | --------- || DiskANN | 1M docs | xms | 0.95 |-| HNSW | 1M docs | xms | 0.95 |+| HNSW | 1M docs | xms | 0.95 | -->*Benchmarks run on Apple M3 Pro 36 GB*@@ -406,14 +401,6 @@ The script will print the recall and search time for each query, followed by the<img src="asset/arch.png" alt="LEANN Architecture" width="800"></p>--### Key Components--1. **🧠 Embedding Engine**: Real-time transformer inference with caching-2. **📊 Graph Index**: Memory-efficient navigation structures-3. **🔄 Search Coordinator**: Orchestrates embedding + graph search-4. **⚡ Backend Adapters**: Pluggable algorithm implementations-## 🔬 PaperIf you find Leann useful, please cite: