WeMM-Embedding

Why we included this project

Retrieval over a mix of screenshots, product photos, scanned pages, and short video clips usually means juggling separate encoders per content type. WeMM-Embedding maps all of those into one shared vector space, so a query about a picture inside a report can match the image, the surrounding text, or an interleaved layout without separate machinery. The repo ships inference examples for both Transformers and Sentence Transformers, plus serving paths through vLLM and SGLang, which keeps the integration work small if you already run an embedding stack. Matryoshka dimensions let you trade storage and latency against accuracy per deployment instead of committing to one fixed vector size. Benchmark results are strong, though audio is not supported and the largest checkpoints run to several gigabytes, so this.