1
00:00:00,180 --> 00:00:04,820
Real shopping is rarely one clean search.

2
00:00:05,180 --> 00:00:09,820
People add constraints, reject choices, and speak vaguely.

3
00:00:10,180 --> 00:00:14,820
That is why we built Shopping Copilot.

4
00:00:15,180 --> 00:00:19,820
The model understands; the system remembers, retrieves, ranks, and explains.

5
00:00:20,180 --> 00:00:24,820
When rule confidence is low, local Qwen interprets intent and action.

6
00:00:25,180 --> 00:00:29,820
It does not generate products, scores, or ranking decisions.

7
00:00:30,180 --> 00:00:34,820
A vague request becomes structured intent.

8
00:00:35,180 --> 00:00:39,820
Item request · Tanks and Camis · adjustable as a soft preference.

9
00:00:40,180 --> 00:00:44,820
Free text stops at the boundary; versioned state records the change.

10
00:00:45,180 --> 00:00:49,820
SQLite FTS5 recalls candidates from fifty thousand frozen products.

11
00:00:50,180 --> 00:00:54,820
Transparent rules rerank the recalled candidates.

12
00:00:55,180 --> 00:00:59,820
The policy asks only the most discriminative follow-up.

13
00:01:00,180 --> 00:01:04,820
Then the shopper changes direction: forget adjustable; use polyester.

14
00:01:05,180 --> 00:01:09,820
The model identifies remove, add, and retain—not isolated keywords.

15
00:01:10,180 --> 00:01:14,820
Version two marks adjustable as superseded and adds polyester.

16
00:01:15,180 --> 00:01:19,820
A bounded rewrite with history—not append-only conflict or reset.

17
00:01:20,180 --> 00:01:24,820
The updated state drives the same retrieval and ranking engine.

18
00:01:25,180 --> 00:01:29,820
Violating products leave; better matches enter; the list reorganizes.

19
00:01:30,180 --> 00:01:34,820
The verified replay places the target at rank one.

20
00:01:35,180 --> 00:01:39,820
Category and material match; the old preference is gone.

21
00:01:40,180 --> 00:01:44,820
The large model sits only at the entrance.

22
00:01:45,180 --> 00:01:49,820
Multi-turn consistency comes from inspectable, versioned state.

23
00:01:50,180 --> 00:01:54,820
Real commerce needs more than task completion.

24
00:01:55,180 --> 00:01:59,820
Sponsored supply, budgets, and relevance require a separate ad layer.

25
00:02:00,180 --> 00:02:04,820
Advertisers configure keywords, bids, and budgets independently.

26
00:02:05,180 --> 00:02:09,820
Those values never enter user state, ranking, or official evaluation.

27
00:02:10,180 --> 00:02:14,820
A relevance floor rejects off-topic inventory—even at a higher bid.

28
00:02:15,180 --> 00:02:19,820
Eligible products compete by bid times relevance; spend is recorded.

29
00:02:20,180 --> 00:02:24,820
The winner enters a clearly labeled sponsored position.

30
00:02:25,180 --> 00:02:29,820
Every organic product remains in exactly the same order.

31
00:02:30,180 --> 00:02:34,820
Public 200: Technical Score 0.8665; private performance remains unknown.

32
00:02:35,180 --> 00:02:39,820
The model and ads are product extensions outside this evaluation contract.

33
00:02:40,180 --> 00:02:44,820
Users experience a conversation—not a metric dashboard.

34
00:02:45,180 --> 00:02:49,820
They can revise, inspect changes, understand ranking, and continue.

35
00:02:50,180 --> 00:02:54,820
The local model may polish replies, but cannot alter facts or ranking.

36
00:02:55,180 --> 00:02:59,820
From vague intent to explainable recommendations and transparent ads.