From 96f74973b1369004a27a22954c8c678b7cffe7c7 Mon Sep 17 00:00:00 2001
From: Andy Lee
Date: Sat, 19 Jul 2025 20:42:52 -0700
Subject: [PATCH] docs: how it works earlier
---
README.md | 25 ++++++++++++-------------
1 file changed, 12 insertions(+), 13 deletions(-)
diff --git a/README.md b/README.md
index c88d10b..c82644d 100755
--- a/README.md
+++ b/README.md
@@ -16,6 +16,7 @@ LEANN is a revolutionary vector database that makes personal AI accessible to ev
RAG your **[emails](#-search-your-entire-life)**, **[browser history](#-time-machine-for-the-web)**, **[WeChat](#-wechat-detective)**, or 60M documents on your laptop, in nearly zero cost. No cloud, no API keys, completely private.
+LEANN achieves this through graph-based selective recomputation with high-degree preserving pruning and dynamic batching, computing embeddings on-demand instead of storing them all. [Read more →](#️-architecture--how-it-works)
## Why LEANN?
@@ -23,7 +24,7 @@ RAG your **[emails](#-search-your-entire-life)**, **[browser history](#-time-mac
-**The numbers speak for themselves:** Index 60 million Wikipedia articles in just 6GB instead of 201GB. Finally, your MacBook can handle enterprise-scale datasets. [See detailed benchmarks below ↓](#benchmarks)
+**The numbers speak for themselves:** Index 60 million Wikipedia articles in just 6GB instead of 201GB. Finally, your MacBook can handle enterprise-scale datasets. [See detailed benchmarks below ↓](#storage-usage-comparison)
## Why This Matters
@@ -217,17 +218,22 @@ This demo showcases how to build a RAG system for PDF/md documents using Leann.
-## How It Works
+## 🏗️ Architecture & How It Works
-LEANN doesn't store embeddings. Instead, it builds a lightweight graph and computes embeddings on-demand during search.
+
+
+
**The magic:** Most vector DBs store every single embedding (expensive). LEANN stores a pruned graph structure (cheap) and recomputes embeddings only when needed (fast).
+**Core techniques:**
+- **Graph-based selective recomputation:** Only compute embeddings for nodes in the search path
+- **High-degree preserving pruning:** Keep important "hub" nodes while removing redundant connections
+- **Dynamic batching:** Efficiently batch embedding computations for GPU utilization
+- **Two-level search:** Smart graph traversal that prioritizes promising nodes
+
**Backends:** DiskANN or HNSW - pick what works for your data size.
-**Performance:** Real-time search on millions of documents.
-
-
## Benchmarks
Run the comparison yourself:
@@ -278,13 +284,6 @@ The evaluation script downloads data automatically on first run.
*Benchmarks run on Apple M3 Pro 36 GB*
-
-## 🏗️ Architecture
-
-
-
-
-
## 🔬 Paper
If you find Leann useful, please cite: