Skip to content

armgit / LEANN

Public Imported from GitHub
0

[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.

docs: add ContextBench agent benchmark (#408)

Jane Hu0784ae628865
diff --git a/README.md b/README.md
index 6f5ae2b..6eb767e 100755
--- a/README.md
+++ b/README.md
@@ -75,6 +75,20 @@ LEANN achieves this through *graph-based selective recomputation* with *high-deg
✨ **No Accuracy Loss:** Maintain the same search quality as heavyweight solutions while using 97% less storage.
+## Making Your Coding Agent Smarter
+
+We plugged LEANN into Claude Code and compared it with BM25 on 30 SWE-Bench Pro tasks from ContextBench, keeping the model, agent, tools, and 8,192-token retrieval budget fixed.
+
+<p align="center">
+ <img src="assets/contextbench_leann_vs_bm25.png" alt="LEANN vs BM25 on ContextBench relevant-code recall, exploration coverage, and agent token usage" width="100%">
+</p>
+
+With LEANN, the agent achieved **2.1× the initial relevant-code recall** (24.2% vs. 11.4%), reached **12.6 percentage points higher relevant-code coverage after exploration** (38.4% vs. 25.8%), and used **8.4% fewer tokens** (3.22M vs. 3.51M).
+
+LEANN surfaces relevant code earlier, so coding agents can spend more of their context window reasoning about the right files instead of repeatedly scanning the repository. **[Reproduce the benchmark →](benchmarks/contextbench/README.md)**
+
+*Note: Results are macro-averaged annotated gold-line metrics over 30 SWE-Bench Pro tasks. Better context access does not guarantee issue resolution.*
+
## Installation
### 📦 Prerequisites: Install uv
diff --git a/assets/contextbench_leann_vs_bm25.png b/assets/contextbench_leann_vs_bm25.png
new file mode 100644
index 0000000..0a55598
Binary files /dev/null and b/assets/contextbench_leann_vs_bm25.png differ