IntentCart — Devpost Submission Copy

Tagline

Stateful, route-aware conversational shopping that recommends and updates context each turn—offline, deterministic, and cost-free at inference time.

Public links

• Source repository: https://github.com/shang-yi-qian/shopping-copilot

How the solution addresses the problem statement

The supplied problem identifies three weaknesses in static keyword search: failure to distinguish Buying from Browsing, failure to accumulate or replace intent across turns, and unnecessary conversational turns. IntentCart addresses each weakness through explicit runtime routing, isolated structured state, and simultaneous ranked retrieval plus proactive clarification.

IntentCart treats conversational shopping as sequential experiment design. Each turn must rank the best products now, expose useful unseen alternatives, and ask the question most likely to reduce uncertainty. This directly matches the challenge's Hit@10, MRR, and mean-turns-to-conversion objectives.

What it does

IntentCart is a headless Python Agent for the organizer's frozen 50,000-product clothing catalog. Every response can return ten ranked parent ASINs and a structured clarification question together.

It handles all four published behaviors:

• Buying: a precision policy locks catalog-supported hard requirements and ranks exact signature matches before broad lexical evidence.
• Browsing: a discovery policy combines category and BM25 recall with a weak supplied-profile prior and conservative product-family coverage.
• Intent Override: selective erasure removes the stale preference, preserves independent evidence, and begins a new recommendation-eligibility epoch.
• Boundary: no-preference answers remain neutral, so the Agent keeps useful recommendations without fabricating a negative constraint.

How we built it

At startup the Agent reads the immutable catalog once and builds:

1. compact product records and deterministic popularity orders;
2. normalized category and catalog-signature reverse indexes; and
3. an in-memory SQLite FTS5 index over text and structured metadata.

reset() deep-copies the organizer's anonymous aggregate profile and creates an isolated state. Every customer-stated constraint records its source, confidence, introduction/confirmation turn, and hard/soft status. Profile tags and summary are normalized into low-confidence terms; they can influence exploration but never filter products and are truncated after two explicit current-session constraints.

At each turn a deterministic orchestrator selects precision, discovery, override_recovery, or fallback. Retrieval merges exact signatures, category, BM25, the permitted profile prior, and deterministic category/global fallbacks. Ranking protects exact evidence, then applies route-appropriate soft signals, popularity, and stable ASIN tie-breaking. V2 promotes catalog popularity ahead of weaker soft tie-breaks only when explicit requirements have narrowed the candidate pool to at most 20 products. Unconstrained Browsing covers no more than two products from one store/title family before deterministic backfill.

When the candidate pool exceeds the result capacity and a useful turn remains, IntentCart asks other while still returning recommendations. Products shown in the current eligibility epoch are not repeated. Exact misses keep the last non-empty pool, FTS errors fall through to catalog order, and missing optional metadata never aborts initialization.

End-to-end demonstration

The repository includes both an executable demonstration and its captured inference evidence:

python3 demo.py

The Browsing trace begins with 519 candidates, asks a clarification while recommending, receives two composition constraints, narrows to one exact candidate, and finds the target at rank 1 on turn 2.

The Intent Override trace shows the target at rank 1 before it is eligible, processes the explicit replacement, increments the eligibility epoch, selects override_recovery, and finds the eligible target at rank 1 on turn 3.

DEMO_OUTPUT.md contains the verified customer turns, safe route traces, ranked predictions, and outcomes. Ground truth is used only after inference to label the result; production Agent code never receives it.

Results

Results below come from the unmodified evaluator over all 200 released public development sessions. They are not private/final-set claims.

Metric	Organizer BM25	IntentCart
Hit@10	0.125000	1.000000
MRR	0.068034	0.856504
MTTC	9.810000	2.090000
Efficiency	0.119000	0.891000
TechnicalScore	0.106710	0.935151

IntentCart achieved 100% Hit@10 in Buying, Browsing, Intent Override, and Boundary. Against immutable submission-v1, v2 preserves that reliability and raises MRR from 0.709905 to 0.856504, improves MTTC from 2.12 to 2.09, and raises TechnicalScore from 0.890571 to 0.935151. All five deterministic sample-ID hash slices improve; these are stability diagnostics, not holdout claims. Two full evaluator runs produced byte-identical JSON. The committed result SHA-256 is 7b553ce517e7c3122a9df21261703027b07e03b0321c1720b60969173065d31e.

On an Apple M2 Pro MacBook Pro with 16 GB memory, macOS 26.5.2, Python 3.12.5, and SQLite 3.45.3:

• Agent initialization: 4.936199 seconds;
• evaluation after setup: 8.873643 seconds;
• cold end-to-end evaluator: 14.49 seconds;
• mean/P95 response latency: 21.130810/40.664084 ms over 418 calls; and
• maximum resident memory: 438,288,384 bytes (418.0 MiB).

Model, API, cost, and feasibility

The selected runtime uses no external model or API. It requires no API key, network call, GPU, vector database, model download, or third-party Python package. Prompt/completion tokens are 0/0 and estimated model cost is US$0.

We tested rather than assumed the value of an LLM. An opt-in OpenAI Responses API path can only reorder the deterministic top-10 set and uses a strict schema, token accounting, a pinned model snapshot, and deterministic fallback. On a bounded 10-session selected-configuration comparison, 11 calls used 25,020 input and 953 output tokens (estimated US$0.006195), two calls fell back safely, and TechnicalScore changed by 0.000000. We therefore disabled it in the frozen default. BM25 is not presented as dense or semantic retrieval.

Impact and relevance

The measured outcome is less search friction in this challenge: the target is covered in all public sessions and found in 2.09 turns on average. In a real catalog, the same state/route pattern could help shoppers explore broadly, converge after giving requirements, and recover gracefully when their need changes. That creates a plausible opportunity for better discovery and conversion, but we do not claim unmeasured revenue lift.

The design is practical for other structured text catalogs because it is deterministic, headless, network-independent, and cheap to run. Real deployment would still require broader language evaluation, canonical product families, live inventory/policy integration, observability, and privacy controls.

Built with

Development tools

• VS Code
• Git and GitHub
• Codex
• macOS terminal tooling

Runtime APIs

• None

Evaluation-only API experiment

• OpenAI Responses API with gpt-5.4-nano-2026-03-17; rejected from the selected runtime after the bounded comparison

Libraries and frameworks

• Python 3.10+ standard library
• SQLite FTS5 through Python's bundled sqlite3
• unittest

Dataset and assets

• Clothing_Shoes_and_Jewelry catalog derived from Amazon Reviews 2023
• 200-session public development set and evaluator
• Original Mermaid architecture diagram and original terminal/demo output
• No product images, storefront captures, logos, webinar clips, third-party footage, or copyrighted music

Limitations and next steps

• The parser is optimized for the organizer's deterministic message templates, not unrestricted customer language.
• BM25/catalog signatures cannot infer every synonym or cross-category use case.
• Profile tags are generic and form only a weak supplied prior; there is no learned identity or cross-session memory.
• Store/title family is an approximation rather than canonical product lineage.
• The optional LLM stage did not improve its bounded comparison and remains disabled in the frozen runtime.
• The public set was fully observed, so no untouched holdout claim is made.

Next we would test a licensed reproducible in-memory dense route, a larger fixed-split semantic-reranking study, more robust language, and information-gain estimation for specific questions. Each would be accepted only after a fixed-split ablation with no scenario reliability loss.

Team contributions

• Tay Kai: Agent engineering and final integration—catalog signatures, indexes, ranking, dialogue state, override handling, adaptive routes, profile conditioning, family coverage, v2 calibration/API experiment, demo tooling, and freeze verification.
• Yi Qian: Evaluation and delivery—evaluator analysis, initial contract/state tests, catalog audits, ablation evidence, reproducibility documentation, and release handoff.
• Together: architecture review, control selection, metrics review, and submission story. Tay Kai completed the enhanced build and remaining artifacts after Yi Qian ended active development.

Built With

• python
• sqlite