ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 8: Windows, overlap and the audio surprise · part 10 of 26
Part 10 · Module 8: Windows, overlap and the audio surprise
Module 8: Windows, overlap and the audio surprise
23 min read·
About 50 minutes · Needs: Module 7 (one embedding space), Module 6 (the test video and its questions) · Code: part2-video-rag-unified/src/extractors/media.py, part2-video-rag-unified/src/ingestion/pipeline.py, part2-video-rag-unified/src/embedders/gemini_embedder.py
Where this fits
This module (highlighted): how Part 2 cuts a video into windows and what exactly it sends to the embedding model. Amber marks a model, blue our own code.
Module 7 chose one multimodal embedding model and one embedding space. This module is about what goes into that model: how long each piece of video is, how the pieces overlap, how they are cut, how their bytes reach Google, and, most of all, what the model really reads. It contains the most important measurement of Part 2: the original v2 believed it embedded picture and speech together, and it did not.
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.