LEANN: 201 GB vs 6 GB?! I Installed This INSANE Vector DB

LEANN says it can index 60 million text chunks in 6 GB instead of 201 GB. That is the README's table, not my test. So I installed it on my Windows laptop, built an index, searched it offline, and asked questions with a local Qwen model. Here are my real numbers, including the crashes and the misses. What I measured (RTX 4060 laptop, CPU torch, one novel, 1,776 chunks): - Install with uv: 4 min 48 s - First build crashed (cp1252 unicode error), fixed with PYTHONUTF8=1 - Graph file 482,681 bytes vs 2,727,936 bytes for raw embeddings = 82% smaller on this book (README claims 97% at 60M chunks, which I did not test) - Cold CPU searches took 48 to 76 seconds each - Worked fully offline when my internet dropped - 3 questions with a 4B local model: 1 miss, 1 partial, 1 misread. Retrieval was fine, the small model was the weak link. Three questions on one book, not a benchmark. Repo: https://github.com/StarTrail-org/LEANN Paper: https://arxiv.org/abs/2506.08276 Chapters 0:00 The 201 GB vs 6 GB claim 0:19 The story and the pain 0:49 The idea: recompute instead of store 1:16 Install on Windows (and the crash) 1:41 The honest disk number 2:06 Search demo and offline proof 2:32 Ask demo with a local LLM 3:09 Verdict 3:29 Subscribe If you want me to test the Claude Code setup next, or the email index, tell me in the comments. Subscribe to Prompt Engineer for videos where I install things and show the crashes and the real numbers. Get GPUs Runpod: https://get.runpod.io/pe48 Get Hostinger: https://www.hostg.xyz/SHJiu Coupon Code "PROMPT" 10% discount 🔗 Connect with me: Email → [email protected] Ko-fi → https://ko-fi.com/promptengineer Patreon → https://www.patreon.com/PromptEngineer975 Book a Call → https://calendly.com/prompt-engineer48/call GitHub → https://github.com/PromptEngineer48 Twitter/X → https://twitter.com/prompt48 Website → https://promptengineer48.com/