ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 20: Evaluating the agent · part 22 of 26
Part 22 · Module 20: Evaluating the agent
Module 20: Evaluating the agent
24 min read·
About 55 minutes · Needs: Module 10 (hit@k, MRR and nearest-rank percentiles), Module 14 (the planner), Module 15 (the cap of 3), Module 16 (product search), Module 17 (grounding) · Code: part3-styleme/backend/evals/run_evals.py, part3-styleme/backend/evals/datasets/, part3-styleme/backend/evals/results/
Where this fits
This module (blue): the agent evals, which drive the planner and the implementer through the real app and re-run the grounding check (green) on every reply.
Modules 14 to 17 built an agent that decides intent before retrieval, asks at most three questions, searches a joint image and title index, and checks its own replies. Each module showed examples that looked right. This module asks the question the whole course is about: is looking right evidence? The answer is a suite of designed tests, run through the real HTTP app, scored by plain arithmetic, and compared with thresholds that each carry a written reason. One of those gates fails, and this module explains why that is the honest result.
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.