Product Suggestor - Real-Time E-Commerce Decision Intelligence

Product Suggestor is an MVP that helps buyers make better purchase decisions before checkout. It analyzes Amazon and Flipkart product links, matches products against a pre-indexed review narrative dataset, and falls back to local LLM analysis when no strong dataset match exists.

Problem Statement

E-commerce shoppers lack reliable decision support before purchase. They frequently face:

• Inconsistent sizing and quality expectations
• Misleading product representation
• Untrustworthy or low-signal review sections

This uncertainty contributes to high return rates (often 20-40%), logistics waste, and reduced user trust.

What This Project Builds

This project is a real-time decision intelligence platform that intercepts buyer uncertainty at the point of consideration.

• Input: A product URL from Amazon.in/Amazon.com or Flipkart.com
• Process: Scrape key product metadata, attempt dataset-based confidence match, then optionally run AI analysis
• Output: A concise objective insight (summary, pros/cons, verdict) to support purchase decisions

Core Approach

The backend uses a two-stage inference strategy:

1. Fast deterministic dataset match
2. AI fallback for unmatched products

Stage 1: Dataset Match (Primary Path)

• dataset/final_narratives.csv is loaded at startup.
• Product titles are indexed with Fuse.js fuzzy search.
• If a confident match is found (score <= 0.35), the API returns precomputed narrative output immediately.

Benefits:

• Low latency
• Consistent output format
• No LLM dependency for known products

Stage 2: AI Fallback (Secondary Path)

• If no confident dataset match exists, the system scrapes on-page reviews.
• Reviews and product metadata are sent to a local Ollama model.
• The model is prompted to return: SUMMARY, PROS, CONS, VERDICT.
• Response text is parsed into structured JSON for the UI.

Architecture

Backend

• Runtime: Node.js + Express
• API route: POST /api/analyze
• Health route: GET /api/health
• Scraping: Puppeteer
• Matching: Fuse.js
• AI: Ollama local inference API

Frontend

• Vanilla HTML/CSS/JS single-page interface
• URL validation for supported domains
• Loading skeleton + rotating status messages
• Result card with source labels:
  • Dataset Match
  • AI Generated
  • Fallback

Request Flow

1. User submits a product URL.
2. API validates domain support.
3. Scraper extracts title, price, rating, review count, image.
4. Matcher searches dataset for confident title match.
5. If matched: return precomputed narrative insight.
6. If unmatched: scrape reviews and generate AI insight via Ollama.
7. UI renders product facts + insight source and recommendation.

Tech Stack

• Node.js
• Express
• Puppeteer
• Fuse.js
• csv-parser
• dotenv
• Ollama (local model serving)

Project Structure

.
├─ server.js                  # App bootstrap, dataset init, route mounting
├─ src/
│  ├─ routes/analyze.js       # Main analyze endpoint
│  ├─ scraper.js              # Amazon/Flipkart extraction logic
│  ├─ matcher.js              # CSV indexing + fuzzy match logic
│  └─ ai.js                   # Ollama prompting + response parsing
├─ public/
│  ├─ index.html              # UI layout
│  ├─ style.css               # UI styling and animation
│  └─ app.js                  # UI behavior + API calls
└─ dataset/
   ├─ final_narratives.csv    # Runtime matching dataset
   └─ complete_results.csv    # Extended processing artifact

Prerequisites

• Node.js 18+
• npm
• (Optional but recommended) Ollama for AI fallback mode

Setup

npm install

Create an .env file in the project root:

PORT=3000
CSV_PATH=./dataset/final_narratives.csv
OLLAMA_URL=http://localhost:11434
OLLAMA_MODEL=gemma2-local

Notes:

• CSV_PATH defaults to ./dataset/final_narratives.csv if unset.
• If Ollama is offline, dataset match mode still works.

Run

Development:

npm run dev

Production:

npm start

Open in browser:

• http://localhost:3000

API

POST /api/analyze

Request body:

{
  "url": "https://www.amazon.in/..."
}

Success response shape:

{
  "product": {
    "title": "string",
    "price": "string",
    "rating": "string",
    "reviewCount": "string",
    "image": "string|null",
    "url": "string",
    "site": "amazon|flipkart"
  },
  "insight": {
    "summary": "string",
    "pros": ["string"],
    "cons": ["string"],
    "verdict": "string",
    "source": "dataset|ai|fallback"
  }
}

Possible error codes:

• 400: Missing/invalid URL or unsupported domain
• 422: Product page scraping failed
• 500: Unexpected server-side error

GET /api/health

Returns:

• Server status and uptime
• Number of indexed products
• Ollama availability and discovered models

Current Limitations

• Supports only Amazon and Flipkart domains.
• Scraping selectors may break if site markup changes.
• Dataset quality can affect narrative reliability.
• AI fallback quality depends on local model availability and hardware performance.
• No automated test suite is currently configured.

Why This Matters

The system demonstrates a practical hybrid design for buyer decision support:

• Deterministic retrieval for speed and consistency
• Generative reasoning for long-tail products
• Source-aware transparency so users know where insight came from

This pattern can reduce uncertainty before purchase, improve trust, and potentially reduce avoidable returns.

Future Improvements

• Add richer trust scoring and confidence calibration
• Expand site support and selector resilience
• Add review authenticity and anomaly checks
• Introduce evaluation benchmarks and automated tests
• Add persistent analytics for decision outcomes