Feedback Analyser

A retail customer feedback dashboard that uses Google Gemini to automatically tag sentiment, categorize each comment, and suggest an action — no manual tagging required.

Screenshots

Dashboard

Submitting new feedback

Overview

Feedback Analyser reads customer feedback (customer, department, comment) and enriches every entry with AI-generated sentiment, category, and a concrete actionable insight, then shows it all on a dashboard with charts and a searchable feedback list. It's a small, self-contained Spring Boot app — feedback is stored in a plain text file rather than a database, so there's nothing to provision before you run it. You can also submit new feedback straight from the UI and watch Gemini analyse it in real time.

Features

• Dashboard with live charts — sentiment, category, and department distributions rendered with Chart.js, with sentiment colors fixed (green/red/purple) so they read consistently no matter how the data is ordered.
• AI-tagged feedback table — every row shows sentiment, category, and an actionable insight, each clearly labeled "AI Generated" so it's obvious what's model output vs. raw data.
• Full feedback list — a dedicated page listing every entry, most recent first.
• New feedback form — submit a customer, pick a department from a dropdown, write a comment, and Gemini analyses it on save.
• Structured Gemini output, not regex scraping — the Gemini request is built with a JSON Schema (ResponseFormat + GenerationConfig) so the model's response is constrained to exactly the fields the app needs, parsed straight into a FeedbackAnalysis record with Jackson.
• Graceful degradation — if the Gemini call fails or returns something unexpected, the app falls back to a clear "Uncategorized" state instead of crashing the page.
• No database required — feedback is persisted to a flat text file and parsed back into objects on read.
• Unit tested Gemini client — GeminiServiceTest and InteractionResponseTest cover the request-building and response-parsing logic with Mockito.

How the AI integration works

Instead of prompting Gemini for free text and hoping it comes back as parseable JSON, this app sends a JSON Schema alongside the prompt (via the Gemini Interactions API's response_format), constraining the model to return exactly sentiment, category, and actionableInsight — nothing more, nothing malformed:

That response is deserialized directly into a Java record with Jackson, so there's no brittle string parsing in the middle.

Architecture

flowchart LR
    Browser -->|Thymeleaf pages| Controller[FeedbackController]
    Controller --> FeedbackService
    FeedbackService -->|read/write| FileService[FileService + ParserService]
    FileService -->|text file| Storage[(sentiment_feedback.txt)]
    FeedbackService --> AnalysisService
    AnalysisService --> GeminiService
    GeminiService -->|Feign client| GeminiAPI[Gemini Interactions API]

Loading

Quick Start

Prerequisites: Java 21, and a Gemini API key.

# set your Gemini API key
export GEMINI_API_KEY=your-key-here

# run the app
./mvnw spring-boot:run

# run the tests
./mvnw test

Then open http://localhost:8080.