LEANN: Real Features vs Marketing
Most vectorDBs are too heavy for personal use -- storing embeddings for millions of documents takes hundreds of GBs.
On the other hand, LEANN takes a fraction of the space -- no cloud or OpenAI dependency.
How it works:
- Instead of storing every embedding, it uses a graph-based selective recomputation strategy, so embeddings are computed on-demand.
- That means you can index 60M text chunks in 6GB instead of 201GB.
Storage savings: 97%, with no accuracy loss.
- Works across your filesystem, emails, browser history, chat logs, and codebases.
- Integrates directly with Claude Code for semantic search in your workflow.
- 100% local, private, and free with Ollama.
The promise: true offline-first, lightweight RAG, where your personal AI knowledge base fits in your pocket.
This post is mostly accurate, but it presents benchmark results and marketing claims as universally applicable facts.
Fact check
"Python library for the 'smallest vector index in the world.'"
🟡 Marketing claim.
LEANN brands itself as "the smallest vector index in the world." While it achieves extremely small index sizes compared with conventional vector databases, this slogan is promotional rather than an independently verified industry fact. (GitHub)
"Takes 97% less storage."
✅ Mostly accurate.
LEANN's published benchmarks show approximately 97% storage reduction compared with traditional vector indexes by storing a compact graph and recomputing embeddings during search. However, this figure comes from the project's own benchmarks and depends on the dataset and backend used. (GitHub)
"Most vectorDBs are too heavy for personal use."
🟡 Opinion with some truth.
Large vector indexes can consume hundreds of gigabytes for very large datasets, making local deployment challenging. However, many modern vector databases also support compression, quantization, and small-scale local deployments, so this statement is an overgeneralization. (arXiv)
"Instead of storing every embedding, it uses a graph-based selective recomputation strategy."
✅ Accurate.
This is LEANN's core innovation. It stores a compact graph structure and recomputes embeddings on demand during retrieval instead of storing every embedding permanently. (arXiv)
"You can index 60M text chunks in 6GB instead of 201GB."
✅ Mostly accurate.
These numbers come directly from LEANN's published benchmark examples. They demonstrate the project's storage savings under a specific benchmark configuration rather than guaranteeing identical results for every workload. (GitHub)
"Storage savings: 97%, with no accuracy loss."
🟡 Needs nuance.
The authors report maintaining retrieval quality while dramatically reducing storage. However, "no accuracy loss" reflects their benchmark results under tested conditions and should not be interpreted as a universal guarantee across every dataset and workload. (arXiv)
"Works across your filesystem, emails, browser history, chat logs, and codebases."
✅ Mostly accurate.
LEANN includes integrations for documents, codebases, Apple Mail, browser history, chat histories, and additional data sources through Model Context Protocol (MCP). Supported sources vary by platform and configuration. (GitHub)
"Integrates directly with Claude Code for semantic search."
✅ Accurate.
LEANN provides an MCP-compatible semantic search service that works with Claude Code, enabling semantic retrieval within coding workflows. (GitHub)
"100% local, private, and free with Ollama."
🟡 Mostly accurate.
LEANN supports fully local deployments using Ollama and local embedding models, allowing private offline operation. However, it also supports cloud providers such as OpenAI and Anthropic, so local operation is optional rather than mandatory. (GitHub)
"The promise: true offline-first, lightweight RAG, where your personal AI knowledge base fits in your pocket."
🟡 Marketing language.
LEANN is designed for lightweight, offline-first RAG, but "fits in your pocket" is a metaphor rather than a measurable technical claim.
Overall verdict
Accuracy: 9/10
The post accurately describes LEANN's core innovation—graph-based selective recomputation to dramatically reduce storage requirements for local RAG. The weaker parts are the universal wording around "no accuracy loss," the marketing slogan "smallest vector index in the world," and the implication that all vector databases are too heavy for personal use.
Real vs Fluff
🟢 90% Real | 🟡 10% Fluff
Real (90%)
LEANN is an open-source Python library.
Uses graph-based selective recomputation.
Achieves significant storage reduction in published benchmarks.
Supports local RAG workflows.
Integrates with Claude Code via MCP.
Works with Ollama and local embedding models.
Supports documents, code, emails, browser history, and other personal data sources.
Fluff / Needs correction (10%)
"Smallest vector index in the world" is marketing, not an independently verified title.
"97% less storage" is benchmark-dependent.
"No accuracy loss" reflects published evaluations rather than a universal guarantee.
"Most vectorDBs are too heavy for personal use" is an overgeneralization. (arXiv)