ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 7: One embedding space with Gemini Embedding 2 · part 9 of 26
Part 9 · Module 7: One embedding space with Gemini Embedding 2
Module 7: One embedding space with Gemini Embedding 2
20 min read·
About 45 minutes · Needs: Module 3 (CLIP and cosine similarity), Module 5 (two collections and rank fusion), Module 6 (the "almost correct" failure) · Code: part2-video-rag-unified/src/embedders/gemini_embedder.py, part2-video-rag-unified/src/core/config.py
Where this fits
This module (highlighted): Part 2's embedding model and the one collection it fills. Amber marks a model, blue our own code.
Part 2 starts here. Module 6 diagnosed the failure: the stitched design finds the right moment for "Which service did Tidewater remove in March?" and still cannot say "Label Printer", because the picture and the speech were ranked in two separate spaces and only their ranks were fused. This module is about the one decision that changes that: a single embedding model that reads a window of video and its speech together and returns one vector. Module 8 covers how the windows are cut and what the model really ingests; Module 9 covers the one collection and the answers.
The rest of this project is yours to keep
This project is bought on its own, once, and stays readable afterwards, including the parts added to it later.