I accidentally built a vector database using video compression

github.com

15 points·saleban1031··3 comments

While building a RAG system, I got frustrated watching my 8GB RAM disappear into a vector database just to search my own PDFs. After burning through $150 in cloud costs, I had a weird thought: what if I encoded my documents into video frames?

The idea sounds absurd - why would you store text in video? But modern video codecs have spent decades optimizing for compression. So I tried converting text into QR codes, then encoding those as video frames, letting H.264/H.265 handle the compression magic.

The results surprised me. 10,000 PDFs compressed down to a 1.4GB video file. Search latency came in around 900ms compared to Pinecone’s 820ms, so about 10% slower. But RAM usage dropped from 8GB+ to just 200MB, and it works completely offline with no API keys or monthly bills.

The technical approach is simple: each document chunk gets encoded into QR codes which become video frames. Video compression handles redundancy between similar documents remarkably well. Search works by decoding relevant frame ranges based on a lightweight index.

You get a vector database that’s just a video file you can copy anywhere.

https://github.com/Olow304/memvid

Now, this is genius, and it has been a long time since I saw something very cool like this. The simplicity and how effective the idea and implementation is is very impressive.

Kudos, I forked it already haha.

Amazing.

Reply on news.ycombinator.com