App store reviews are a goldmine of user feedback, but manually tracking trends across thousands of daily reviews is impossible. Product teams need a way to see "what's trending" (e.g., "Login failing after update") in real-time. This system uses AI agents to process reviews, extract topics, and track their volume over a rolling 30-day window.
The system is built as a pipeline of specialized agents:
- Ingestion: Loads reviews (Simulated for this demo to guarantee trend visibility).
- Topic Extraction Agent: Uses Sentence Embeddings (all-MiniLM-L6-v2) and Clustering to group reviews into semantic topics without needing pre-defined keywords.
- Deduplication Agent: Maps new extracted topics to a canonical "Topic Memory" to ensure "App crashing" and "Crahsed app" are counted as the same trend.
- Evolution Agent: Identifies truly new topics that haven't been seen before and registers them.
- Trend Agent: Aggregates counts and produces the final rolling window report.
The final output is a tabular trend report where rows represent canonical topics and columns represent dates from T-30 to T. Each cell contains the frequency of that topic on the given day. The report is exported as a CSV for easy consumption by product teams.
- Recall over Precision: We prefer to surface a potential issue even if the label is slightly noisy, rather than burying a critical bug.
- Explainability: Every specific review contributes to a topic count. We can trace a trend spike back to individual user quotes.
- No Black Box: We use explicit clustering and similarity thresholds, avoiding opaque end-to-end LLM calls for the core counting loop to ensure speed and low cost.
- Sarcasm: Sentiment is not deeply analyzed; "Great job breaking the app" might be clustered with "App broken".
- Context: Short reviews like "Bad" are filtered or ignored as they don't provide actionable topics.
- Install dependencies:
pip install -r requirements.txt - Run the demo directly:
python demo.py