CheckWise

CheckWise is a multi-agent framework designed for high-precision content forensics and automated text verification. By utilizing a hybrid architecture, the platform combines real-time cloud-based web grounding with deep local statistical analysis to dissect text for credibility, structural integrity, and origin.

Key Capabilities

• Generative & Synthetic Pattern Detection: Uncovers machine-generated structures, tracking rhythmic monotony, repetitive syntax, and typical AI-authored signatures.
• Factual Grounding & Hallucination Auditing: Deploys a team of specialized, coordinated agents—including Fact-Checking, Grammatical, and Statistical units—to cross-reference statements against live web databases, isolating factual hallucinations.
• Smart URL Processing & Extraction: Seamlessly ingests and sanitizes raw articles, reports, or text files directly via URL extraction to deliver immediate, multi-layered credibility scores and highly detailed reasoning analytics.

App demo

Requirements

• Backend: Python 3.12+ (packages in requirements-statistical-agent.txt), SQLite (checkwise.db).
• Frontend: Node.js v20+ & npm (React with Vite).
• Testing: Playwright (requires npx playwright install --with-deps).

Quick Start

Create a virtual environment, activate it, and install the required dependencies:

# Create virtual environment
python -m venv venv

# Activate on Windows (CMD)
call venv\Scripts\activate
# Activate on Windows (PowerShell)
# .\venv\Scripts\Activate.ps1
# Activate on macOS/Linux
# source venv/bin/activate

# Install dependencies
pip install --upgrade pip
pip install -r requirements-statistical-agent.txt

Install the Node.js packages in a separate terminal:

npm install

Start both services in separate terminal sessions:

# Ensure venv is active
python -m uvicorn backend.app:app --reload --port 8000

npm run dev

Describing the app - (non-tehnical description - user stories, backlog creation, features list, design description, behavior description)

1. User stories

https://docs.google.com/document/d/1O0QkMZ_vKr_x_rvjXHyJ5c-q4CBUVyEVc0pRYqqXhHk/edit?tab=t.0

2. Backlog Creation

We monitorized our backlog creation using Trello. It can be found here.

3. Features list

• Automated URL Text Extraction: Ingest and sanitize raw text from any webpage using high-performance scraping.
• Multi-Agent Evaluation: Coordinate specialized Fact-Checking, Grammatical, and Statistical agents to evaluate credibility.
• AI Content Forensics: Analyze linguistic patterns, syntax, and rhythmic monotony to identify synthetic (machine-generated) text.
• Factual Hallucination Auditing: Cross-reference extracted claims with live data to flag discrepancies and hallucinations.
• Multi-Layered Scoring System: Generate an intuitive, overall credibility score backed by granular analytical details.
• Interactive Dashboard: Seamlessly track verification history, visualize agent feedback, and review detailed reasoning.
• Comprehensive Testing Pipeline: Automated end-to-end user flow testing with Playwright alongside backend verification via Pytest.

4. App's behaviour description

Basically, the user can either opt for pasting raw text or providing a URL to be verified. If they choose to analyze a URL, the system will use a high-performance web scraper to extract and clean the text content from the source webpage. Either way, the platform dispatches the text into a multi-agent evaluation pipeline.

Within this pipeline, specialized coordinate agents—specifically the Fact-Checking Agent, the Grammatical Agent, and the Statistical Agent—work concurrently. The statistical engine detects synthetic patterns, repetitive syntax, and rhythmic monotony to identify if the text is AI-generated, while the fact-checking engine cross-references claims against live web databases to spot factual hallucinations.

Once the analysis is finalized, an overall credibility score is dynamically generated and displayed on the screen. Alongside this score, the user is presented with a detailed, interactive report containing granular agent breakdowns, highlighted text snippets, and reasoning analytics. The user can then explore the flags, inspect specific warnings, or clear the current analysis to start a new verification session.

5. UML Diagrams, Workflows, Components Arhitecture

It can be found here

Source control

Branches: https://github.com/adapreda/CheckWise/branches

• branch branch-ada - contains statistical agent;
  
• branch alina - contains database remastering and switch to ollama3.1:8b;
  
• branch marius - contains fact-checing agent;
  
• branch fullagent - contains fact-checking agent clone;
  
• branch branch_2 - contains code that allows the user to select an image from the gallery or to take a picture which will be scanned. The scanner is programmed to extract the name of the product and its price;
  
• branch url-processing - contains the code for URL extraction and processing;
  
• branch fabi - contains Gramatical Agent, UML diagrams, and AI usage report
  

Commits: https://github.com/adapreda/CheckWise/commits

CI/CD Pipeline and Automated tests

The project features an automated testing and continuous integration/continuous deployment (CI/CD) system configured using GitHub Actions.

1. Automated Testing

• Backend (Pytest): Unit tests (e.g., backend/tests/test_health.py) verify API health using TestClient. Run with pytest backend/tests.
• Frontend (Playwright): Test (e.g., tests/login.spec.ts) simulate a real user accessing http://localhost:8080/login, entering credentials, and verifying redirection to /checker.

2. GitHub Actions (Pipeline)

Configuration is managed via files in .github/workflows/:

• CI (ci1.yml): Runs automatically on every push or pull_request to main/develop. Installs dependencies, runs pytest and Playwright to prevent bugs.
• CD (cd.yml): Runs exclusively on push to main. Validates the code, builds the final production version (npm run build), and generates an artifact (application-release) containing everything needed for deployment.

3.Local Execution

Before pushing, make sure tests pass locally:

# Backend
pytest backend/tests

# Frontend E2E
npx playwright test tests/login.spec.ts