Maybe we need consider intergration MP4 as vector storage

[spring-ai-tech]

introduction

Specifically, it employs an ingenious encoding scheme: each block of text is converted into a QR code image, and these images are assembled as frames in a video. By leveraging MP4’s video compression algorithms, this approach achieves a compression ratio up to 10 times higher than traditional text storage methods. Meanwhile, the system generates a companion JSON index file that records the position of each text block within the video.

The most innovative part lies in its semantic search capability: it uses sentence-transformers to generate text embeddings and FAISS for similarity search. From a technical architecture standpoint, Memvid is essentially a hybrid system—video files are used to store the raw data, while a vector index powers efficient retrieval. In effect, this means you can fit an entire library into a single MP4 file and instantly locate any piece of information using natural language.

具体来说，它采用了一种巧妙的编码方案：每个文本块被转换成QR码图像，这些图像成为视频的帧。通过利用MP4的视频压缩算法，可以实现比传统文本存储高10倍的压缩率。同时，系统会生成一个配套的JSON索引文件，记录每个文本块在视频中的位置信息。最关键的是，它使用sentence-transformers生成文本嵌入，通过FAISS进行相似度计算，实现语义搜索功能。从技术架构来看，Memvid实际上是一个混合系统：视频文件负责存储原始数据，而向量索引负责实现快速检索。这意味着，你可以把整个图书馆装进一个MP4文件里，然后用自然语言瞬间找到任何信息。

https://github.com/Olow304/memvid