ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 9: One collection, no fusion, grounded answers · part 11 of 26
Part 11 · Module 9: One collection, no fusion, grounded answers
Module 9: One collection, no fusion, grounded answers
22 min read·
About 45 minutes · Needs: Module 5 (Qdrant points and rank fusion), Module 7 (one embedding space), Module 8 (windows and the audio part) · Code: part2-video-rag-unified/src/retrieval/retriever.py, part2-video-rag-unified/src/generation/generator.py, part2-video-rag-unified/src/vectorstore/qdrant_store.py
Where this fits
This module (highlighted): Part 2's one collection, its one search per question and the answer built from notes. Amber marks a model, blue our own code.
Modules 7 and 8 produced one Gemini Embedding vector per window. This module follows a question through the rest of Part 2: one search in one collection, five windows back, and an answer that names what was on screen and cites when. It also shows the second change Part 2 made, which is easy to miss and important to state honestly: at ingest a generation model writes notes about every window, and the answers are written from those notes.
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.