ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 6: "Almost correct": diagnosing the failure · part 8 of 26
Part 8 · Module 6: "Almost correct": diagnosing the failure
Module 6: "Almost correct": diagnosing the failure
22 min read·
About 50 minutes · Needs: Module 1 (the problem with multimodal search), Module 5 (two collections and rank fusion) · Code: samples/build_test_video.py, samples/questions.jsonl, part1-video-rag-stitched/src/generation/generator.py
Where this fits
This module: the test video and its questions, the eval that scores Part 1, and the path from rank fusion to the answer model (amber, a model).
Part 1 is built and hardened. This module asks whether it works, and answers with a designed test instead of a few happy questions. The highlighted pieces are the test video and its 26 questions, the eval that scores Part 1 through its API, and the last hop of the query path: from rank fusion to the answer model. The eval harness itself is Module 10; here we read its results for Part 1 and find the cause.
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.