ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 5: Two collections and rank fusion · part 7 of 26
Part 7 · Module 5: Two collections and rank fusion
Module 5: Two collections and rank fusion
19 min read·
About 45 minutes · Needs: Module 3 (frames and CLIP vectors), Module 4 (transcript chunks and MiniLM) · Code: part1-video-rag-stitched/src/vectorstore/qdrant_store.py, part1-video-rag-stitched/src/retrieval/fusion.py, part1-video-rag-stitched/src/retrieval/retriever.py
Where this fits
This module: the two Qdrant collections and the rank fusion that merges their results (blue); CLIP and MiniLM (amber) make the two query vectors.
Modules 3 and 4 turned the test video into 40 frame vectors and 12 transcript vectors. This module is where they are stored and searched: two Qdrant collections, one per vector space, and a 12-line function that merges the two result lists by rank. That merge is the only place in Part 1 where picture and speech meet, which makes it the heart of the design and of its failure.
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.