How to use Codex to preserve and reset shopping filters
Codex-assisted development produced a shopping state machine that retains useful constraints and forgets replaced intent; gains concern public evaluation.
devpost.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Improve this deterministic shopping baseline using a frozen catalog and held-out evaluation. Separate active constraints from history, exclude rejected products, and reset conflicts when intent changes. Test lexical and semantic retrieval independently, validate exact product IDs and report development scores separately from generalization or live purchase claims.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Wen Yi Liang, Rayhan Naufal and Qingya He
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The source reports public-development Hit Rate@10 improving from 0.125 to 0.870 on 200 sessions; no private-evaluation or actual purchasing outcome is established.
- Human involvement
- The team chose evaluation objectives, froze configurations and rejected failed variants; public development results were kept separate from unseen users and live purchases.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- The team examined failures of a message-by-message BM25 shopping baseline with Codex-assisted development.
- It implemented active intent, exclusions and controlled forgetting, then tested retrieval and clarification components independently.
- Humans froze configurations and used a deterministic evaluator and central result validator before accepting changes.
Wen Yi Liang’s team used Codex-assisted development to build Your Shopping Bestie, a local shopping-retrieval agent evaluated against a frozen product catalog. Its specific problem was remembering useful requirements while dropping preferences the shopper had replaced.
The starting BM25 agent searched each message independently. A rejection such as not quite right became a useless query, losing the original product constraints. The revised state machine retained the last informative request, excluded previously shown products and separated active requirements from historical hints. A changed product category reset conflicting state rather than carrying obsolete shoe preferences into a jacket search.
The team tested lexical, dense and hybrid retrieval, finding that plausible semantic matches did not necessarily contain the evaluator’s exact product identifier. It favored catalog and lexical evidence, enforced hard constraints and checked valid identifiers, duplicates and result count centrally.
The source reports Hit Rate@10 rising from 0.125 to 0.870 across 200 public development sessions. Those are development-set results, not private-test generalization or actual purchase success. Humans froze configurations, changed one component at a time and rejected variants that failed the gate. Codex assisted the build; the final runtime was deterministic and used no hosted model inference. The case is an executed shopping-agent evaluation, not a live-price recommendation.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Devpost · Wen Yi Liang, Rayhan Naufal and Qingya He
Frequently asked questions
Does the final agent require a hosted model?
No. The source describes a fully local, deterministic runtime.
How does it handle vague rejection?
It reuses the last informative request, retains valid constraints and excludes products already shown.
Do the reported scores prove unseen-user success?
No. They are results on the public development sessions; broader generalization remains future work.
More details
About this account
The team examined failures of a message-by-message BM25 shopping baseline with Codex-assisted development. It implemented active intent, exclusions and controlled forgetting, then tested retrieval and clarification components independently.
Firsthand dated evaluation opened; source explicitly says Codex-assisted development. Final runtime is deterministic, not Codex or an LLM. Frozen-catalog public-development metrics do not establish current prices, private-test performance or purchases. Suggested prompt written by Agents In Use; no original task prompt was found in the verified text.
Credit: Wen Yi Liang, Rayhan Naufal and Qingya He — Devpost
Source published
Added to Agents In Use
Source checked
Tools and topics
Codex · Python · BM25 · TechJam evaluator · frozen product catalog
changed shopping intent · catalog retrieval · controlled evaluation
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.