ProjectIntent Over Keywords: From Stitched Embeddings to a Production Fashion Agent · Module 10: Measuring v1 against v2 · part 12 of 26
Part 12 · Module 10: Measuring v1 against v2

Module 10: Measuring v1 against v2

22 min read·
About 50 minutes · Needs: Module 6 (the test video and its question kinds), Module 8 (windows and saturation), Module 9 (notes and answers) · Code: evals/video/run.py, evals/video/reports/20261027T192400Z.json

Where this fits

WHERE THIS FITS Intent Over Keywords this module PART 1: STITCHED VIDEO RAG three models, two spaces, RRF PART 2: UNIFIED VIDEO RAG one model, one space PART 3: STYLEME FastAPI, LangGraph, React QUALITY AND OPERATIONS evals, Docker, Cloud Run Frames ffmpeg, one per 5 s CLIP ViT-B/32 frames + text, 512 d Whisper base speech to text MiniLM 64-word chunks, 384 d Two collections p1_visual, p1_text Rank fusion RRF, k = 60 Gemini answer from transcript text API + Streamlit upload, ask, sources Windows ffmpeg, video + audio Embedding 2 Gemini, one vector One collection p2_unified, no fusion Notes + answer Gemini, cited times React web chat, photo, chips nginx one URL, /api proxy FastAPI /api/chat, limits Sessions in memory, cap of 3 Planner intent before retrieval Implementer batched embed, search Generator stylist reply, ids Grounding check no LLM, drops invented Gemini flash-lite, embedding 2 Qdrant styleme_products Catalogue ingest image + title vectors Langfuse optional traces Test video 26 questions, 4 kinds Video eval Part 1 vs Part 2 Agent evals intent, grounding Guards uploads, limits, secrets Containers compose profiles Cloud Run StyleMe only CI/CD GitHub Actions, WIF Original code v1, v2, fashion-agent
This module (highlighted): the harness that runs the test video's 26 questions through both parts and scores them. Amber marks a model, blue our own code.

Part 1 ends with a design that looks right and answers "almost correct". Part 2 claims to fix it. This module is the evidence: a harness that asks both parts the same 26 questions about the same video, through the same API a user would call, and turns the replies into numbers. It also says plainly what those numbers prove and what they do not, because an answer that looks right is not evidence; a designed test is.

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.

Buy this project · ₹799See what is inside

Or open everything with a plan