IntentCompass

Inspiration

Conversational shopping is difficult because intent is incomplete, changes over time, and often contains “no preference” answers. A search system that simply appends every message can preserve stale constraints and miss the exact product. IntentCompass treats active intent as explicit, replaceable state and changes retrieval strategy as the conversation evolves.

What it does

IntentCompass is a local-first Track 4 Python agent that returns an ordered Top 10 from the supplied 50,000-product catalog on every turn. It supports Buying/Browsing routing, intent override, three-valued constraints, multi-route lexical recall, on-demand dense retrieval, bounded diversity, guarded local cross-encoder ranking, proactive clarification, recovery after rejection, and an explicit consent-based profile handoff.

How we built it

The official Agent.reset/Agent.respond adapter delegates to modular session, workflow, retrieval, and ranking code. One in-memory SQLite FTS5 index serves keyword, category, exact-constraint, and expansion routes. Pinned ONNX MiniLM models provide optional in-memory dense recall and local cross-encoder ranking only when a request qualifies. Missing model assets fail closed to the lexical path. The evaluator, data, metrics, and stopping rules remain unchanged.

Challenges we ran into

The hardest problems were replacing stale intent without leaking evaluator targets, balancing exact-item ranking against catalog coverage, and keeping optional semantic components deterministic and lightweight across Windows, macOS, and Linux. We also had to distinguish genuine improvements from Public-set overfitting, so every candidate passed reproducible gates and negative experiments were retained rather than forced into the release.

Accomplishments that we're proud of

On all 200 released Public sessions, IntentCompass reproduced HitRate@10 0.98, MRR 0.696861, and MTTC 3.755, for TechnicalScore 0.843958. TechnicalScore is an input to Technical Execution, not the overall competition result. Our demo shows an Intent Override session in which stale intent is removed on turn 4 and the exact target first appears on turn 5 at rank 8. The target is visible only to the evaluator/demo harness, never to the Agent.

What we learned

We learned that explicit conversational state, deterministic fallbacks, and bounded route activation can outperform adding complexity everywhere. We also learned to treat HR, MRR, and MTTC as a Pareto problem: a higher recall candidate is not automatically better if it lowers precision or adds unnecessary turns.

Built with

Python 3.12/3.13 and SQLite FTS5
NumPy, ONNX Runtime, and Hugging Face Tokenizers
all-MiniLM-L6-v2 dense encoder and ms-marco-MiniLM-L6-v2 cross-encoder
Optional Qwen/DashScope and DeepSeek adapters with bounded cost and fallback
Git, GitHub Actions, and AI-assisted coding tools for implementation and QA
Organizer-provided Amazon Reviews 2023-derived text catalog and sessions

Feasibility, cost, and limitations

The main system is CPU-based, text-only, and in-memory; it requires no GPU, training, multimodal processing, or industrial vector database. The Public run attempted no network calls and used zero model tokens. Independent Win11 validation measured about 143 ms p95 with local semantic assets, which is machine-specific. A successful optional Qwen proof cost an estimated RMB 0.058274 but is not claimed as a metric improvement. Public and synthetic tests do not guarantee the organizer's private-set or real business performance.

What's next

We would validate the same target-blind policies on larger private and real-world conversational traffic, add multilingual catalog normalization, and calibrate the semantic route under strict latency and cost budgets.

Team

Fang Tianchen led architecture, algorithms, and integration. Liu Chunyi handled cross-platform QA and reviews. Cheng Xianyun handled analysis, evidence, Windows testing, and video production. Wang Siwen contributed retrieval experiments, benchmark/evidence integrity, and semantic QA. AI coding assistance was used; the team owns the final design and verification.