TrustCart AI

TrustCart AI is a free, local-first Streamlit app for analyzing whether product reviews look trustworthy. It accepts pasted reviews, uploaded CSV files, or bundled sample data, then produces sentiment, emotion, fake/spam risk, product aspects, a weighted trust score, and a final buy/avoid report.

No paid APIs are required. Optional Hugging Face models run locally when available, and each model-backed feature has a deterministic fallback so the app can still run in constrained environments.

Model Pipeline

1. Review loading

   • Paste text, upload CSV, or use data/sample_reviews.csv
   • Minimum review length filtering
   • Optional rating column support

2. Sentiment and emotion

   • Hugging Face transformers pipelines when available
   • Rule-based fallbacks if model loading or downloads fail

3. Fake/spam review detection

   • Linguistic signals: short reviews, repeated wording, uppercase ratio, exclamation marks, generic phrases, rating extremity
   • Sentence similarity with sentence-transformers/all-MiniLM-L6-v2
   • IsolationForest anomaly detection on engineered features
   • Safe fallback to non-embedding features if the sentence model cannot load

4. Aspect mining

   • KeyBERT keyword extraction when available
   • YAKE fallback
   • Keyword-frequency fallback
   • Groups mentions into product aspects such as battery, display, performance, delivery, quality, durability, comfort, support, and price

5. Trust scoring

   • Weighted score from sentiment, aspect quality, rating consistency, review volume, and authenticity
   • Labels: Strong Buy, Buy with Caution, Mixed / Compare Alternatives, Avoid
   • Warnings for high fake risk, duplicate reviews, negative aspect dominance, and too few reviews

6. Final report

   • Verdict
   • Pros and cons
   • Suspicious review patterns
   • Who should buy
   • Who should avoid
   • JSON and CSV export

Tech Stack

• Streamlit
• pandas
• numpy
• scikit-learn
• transformers
• sentence-transformers
• KeyBERT
• YAKE
• Plotly
• joblib
• PyTorch CPU wheel

How To Run

cd trustcart-ai
pip install -r requirements.txt
streamlit run app/streamlit_app.py

If you are running from the repository root:

cd /home/user/Desktop/TrustCart/trustcart-ai
pip install -r requirements.txt
streamlit run app/streamlit_app.py

CSV Format

Uploaded CSV files must include a review column. A rating column is optional.

review,rating
"Great battery life and solid build.",5
"Stopped working after one week.",1

Screenshots

Add screenshots here after running the app locally:

• screenshots/trust-score.png
• screenshots/fake-review-risk.png
• screenshots/final-report.png

Outputs

The app can export the final report from the UI:

• trustcart_report.json
• trustcart_report.csv

It also stores the latest analysis artifact in:

outputs/latest_analysis.joblib

Optional Local Environment

No environment variables are required. See .env.example for optional local model cache settings.

Limitations

• The fake-review detector is heuristic and statistical, not proof of fraud.
• Hugging Face models may need an initial download and can be memory intensive.
• Aspect grouping is designed around common consumer-product categories and may miss niche domain terms.
• Sentiment and emotion models can misread sarcasm, mixed reviews, or multilingual text.
• Trust score weights are transparent defaults, not a certified risk model.

Future Scope

• Add multilingual review support
• Add per-category aspect dictionaries
• Add trend analysis across review dates
• Add reviewer metadata support when available
• Add user-adjustable trust-score weights
• Add automated screenshots and a lightweight test suite