Ask What Matters

Ask What Matters is a submission-ready prototype for the Wharton/Expedia Hack-AI-thon. It uses the provided Expedia property datasets to detect what information is missing or stale for a hotel, asks one smart follow-up question, and captures the traveler’s answer by text or voice.

For a short recorded walkthrough, use https://www.loom.com/share/d5d2de466ea049969134275af875d84b.

Challenge Framing

Static review prompts miss what matters most for a specific property right now. This prototype focuses on one polished flow:

1. A data-driven agent identifies the most useful missing or stale topic for that property.
2. A question-generation agent asks one short follow-up.
3. An integration agent saves the answer and summarizes what new information was learned.

3-Agent Workflow

1. Gap Analysis Agent

• Reads historical reviews from Reviews_PROC.csv
• Reads property context from Description_PROC.csv
• Chooses one target topic such as cleanliness, service, room comfort, or amenities
• Explains why the topic is missing, stale, or otherwise valuable

2. Question Agent

• Turns the selected topic plus current review text into one concise follow-up question
• Uses OpenAI for question generation and audio when configured
• Falls back to local heuristic prompts when OPENAI_API_KEY is not present

3. Integration Agent

• Takes the original review, follow-up question, and traveler answer
• Produces a short “what we learned” summary
• Produces an enriched review snippet and saves it to local demo storage

Data Usage

The prototype uses both provided Expedia datasets:

• data/Reviews_PROC.csv Used to detect topic coverage, missing themes, stale themes, and sentiment signals for a specific property.
• data/Description_PROC.csv Used to provide property context such as location, amenities, and property description to ground the question-generation step.

Topic Taxonomy

The demo uses a compact topic taxonomy:

• cleanliness
• service
• room_comfort
• amenities
• condition
• location
• value
• safety

Recency Heuristic

A topic is considered:

• missing if it never appears in historical reviews for the property
• stale if it appears in older reviews but not within the latest 6 months of review history
• covered otherwise

This heuristic is intentionally simple and explicit for demo clarity.

Quick Start

npm install
cp .env.example .env
npm run dev

Open http://localhost:3000.

Environment

Set the following in .env for full functionality:

OPENAI_API_KEY=sk-...

Optional overrides:

OPENAI_CHAT_MODEL=gpt-4o-mini
OPENAI_TTS_MODEL=tts-1
OPENAI_TTS_VOICE=alloy
OPENAI_TRANSCRIPTION_MODEL=whisper-1
PORT=3000

Fallback Behavior

If OPENAI_API_KEY is missing:

• question generation still works using local heuristics
• the UI clearly indicates fallback mode
• audio playback is disabled
• transcription returns a clear message instead of failing silently
• integration still produces a local saved summary

API Summary

• GET /api/health Returns backend status, OpenAI configuration status, and demo property context.
• POST /api/generate-question Runs the gap-analysis agent and question agent.
• POST /api/transcribe Transcribes recorded voice input with OpenAI Whisper.
• GET /api/answers Returns saved answer history for the selected property.
• POST /api/answers Runs the integration agent and saves the answer record.

Local Demo Notes

• npm run dev starts the Express server and serves the static frontend.
• The demo property defaults to Expedia property id db38b19b897dbece3e34919c662b3fd66d23b615395d11fb69264dd3a9b17723.
• Saved answers are stored in data/savedAnswers.json for demo persistence.

Deployment

The hackathon submission requires a public prototype link. That means:

• for development, rehearsal, and video recording, npm run dev is enough
• for the final submission form, you should also deploy the app to a public URL

Recommended simple hosts:

• Vercel
• Replit
• Hugging Face Spaces

Minimum deployment checklist:

• set OPENAI_API_KEY in the host environment
• make sure .env is not committed
• confirm the public link loads without any local setup
• test at least one full review -> question -> save flow on the hosted app

Submission Notes

For local recording, npm run dev is enough. For the actual submission, provide the hosted public link in addition to the repository link.

Before submission:

• keep data/savedAnswers.json empty or with only intentional demo records
• verify .env is not committed
• confirm the public link loads without requiring local setup

Repository Structure

• src/server.js - Express server entry point
• src/routes/api.js - API endpoints and 3-agent orchestration
• src/reviewAnalysis.js - compatibility export surface for review intelligence
• src/config.js - OpenAI client configuration
• data/savedAnswers.json - local saved answer history
• docs/ - demo script, architecture notes, and hackathon reference materials

What Is Intentionally Simplified

• One polished property flow instead of broad multi-property productization
• A transparent heuristic topic taxonomy instead of a trained classifier
• Sequential agent orchestration inside one backend rather than a complex multi-agent runtime
• Local JSON persistence instead of a production database