SCORE - SCORE Curates Organizational Repository for Embeddings

Enterprise document repository analysis tool. Ingests documents from multiple sources, detects duplicates, extracts claims, finds contradictions, clusters topics, identifies documentation gaps, flags hallucination risks, runs RAG quality audits, and produces a Nutri-Score-style quality grade (A-E).

Built with Django, SQLite + sqlite-vec, Celery, OpenAI/Azure OpenAI, and spaCy.

<p align="center"> <img src="docs/assets/welcome.png" alt="Welcome" width="45%" /> <img src="docs/assets/dashboard.png" alt="Dashboard" width="45%" /> </p> <p align="center"> <img src="docs/assets/connectors.png" alt="Connectors" width="45%" /> <img src="docs/assets/analysis.png" alt="Analysis" width="45%" /> </p> <p align="center"> <img src="docs/assets/rag_audit.png" alt="RAG Audit" width="45%" /> <img src="docs/assets/chatbot.png" alt="Chatbot" width="45%" /> </p>

---

Table of Contents

• Architecture
• Prerequisites
• Installation
• Docker Deployment
• Configuration
• Running the Application
• Tests
• Dependencies
• Documentation
• Contributing
• License

---

Architecture

┌─────────────┐     ┌──────────────┐     ┌──────────────────┐
│  Dashboard   │────▶│  Django Web  │────▶│  SQLite (ORM)    │
│  (Bootstrap  │     │  Server      │     │  db.sqlite3      │
│   + D3.js)   │     └──────┬───────┘     └──────────────────┘
└─────────────┘            │
                           │ enqueue
                           ▼
                    ┌──────────────┐     ┌──────────────────┐
                    │  Celery      │────▶│  sqlite-vec      │
                    │  Worker      │     │  vec.sqlite3     │
                    └──────┬───────┘     └──────────────────┘
                           │
                    ┌──────┴───────┐
                    │  OpenAI /    │
                    │  Azure API   │
                    └──────────────┘

• Web server: Django 5.1, serves the dashboard, chat interface, and triggers async jobs.
• Task queue: Celery workers run ingestion, analysis, and audit pipelines.
• Primary database: SQLite via Django ORM (data/db.sqlite3) for all relational data.
• Vector database: Separate SQLite file (data/vec.sqlite3) with the sqlite-vec extension for KNN embedding search.
• LLM: Unified client supporting OpenAI, Azure OpenAI, and Azure Mistral for embeddings, chat completions, and JSON-mode structured output. Supports fallback models on rate-limit errors.
• Semantic graph: spaCy-based concept extraction with NetworkX graph and FAISS index for knowledge-map visualizations.

---

Prerequisites

• Python 3.12+
• Redis (for production Celery broker) or use the SQLAlchemy database broker for local development
• An OpenAI API key, Azure OpenAI deployment, or Azure Mistral deployment
• spaCy model (optional, for semantic graph): python -m spacy download fr_core_news_sm

---

Installation

Quick Start

bash scripts/run_dev.sh

This script will:

1. Create a .venv virtual environment
2. Install all dependencies from requirements.txt
3. Copy .env.example to .env (with database broker for dev)
4. Run Django migrations
5. Load sample data (users, tenants, documents)
6. Collect static files

Manual Setup

# Create and activate virtual environment
python3.12 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Copy and edit environment config
cp .env.example .env
# Edit .env with your API keys and settings

# Run migrations
python manage.py migrate --run-syncdb

# Load sample data (optional)
python scripts/load_sample_data.py

# Collect static files
python manage.py collectstatic --noinput

Default Credentials (from sample data)

WARNING: These are development-only credentials. Change them immediately in any non-local environment. Never use default passwords in production.

| User | Password | Role | |-------|----------|-------------------| | admin | admin | Superuser (admin) | | demo | demo | Editor |

---

Docker Deployment

A multi-stage Dockerfile is provided for production deployment.

# Build the image
docker build -t score .

# Run the container
docker run -p 8000:8000 \
  -e SECRET_KEY=your-production-secret-key \
  -e DEBUG=False \
  -e LLM_PROVIDER=openai \
  -e OPENAI_API_KEY=sk-... \
  score

The Docker image:

• Uses a multi-stage build (builder + runtime) for smaller image size
• Runs with gunicorn (4 workers, 4 threads)
• Includes a /healthz/ health check endpoint (30s interval)
• Exposes port 8000

---

Configuration

SCORE reads configuration from two sources, with environment variables taking precedence.

.env (environment variables)

# Django
# IMPORTANT: Generate a real secret key for production:
#   python -c "from django.core.management.utils import get_random_secret_key; print(get_random_secret_key())"
SECRET_KEY=change-me-in-production
DEBUG=True
ALLOWED_HOSTS=localhost,127.0.0.1

# Celery broker
CELERY_BROKER_BACKEND=database    # "database" for dev, "redis" for production
CELERY_BROKER_URL=redis://localhost:6379/0

# LLM Provider: "openai", "azure", or "azure_mistral"
LLM_PROVIDER=azure_mistral

# OpenAI
OPENAI_API_KEY=sk-...
OPENAI_CHAT_MODEL=gpt-4o
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
OPENAI_EMBEDDING_DIMENSIONS=1536

# Azure OpenAI (if LLM_PROVIDER=azure)
AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_API_VERSION=2024-06-01
AZURE_OPENAI_CHAT_DEPLOYMENT=gpt-4o
AZURE_OPENAI_EMBEDDING_DEPLOYMENT=text-embedding-3-small
# Optional: separate Azure endpoint for embeddings (if different resource)
AZURE_OPENAI_EMBEDDING_ENDPOINT=
AZURE_OPENAI_EMBEDDING_API_KEY=

# Azure Mistral (if LLM_PROVIDER=azure_mistral)
AZURE_MISTRAL_ENDPOINT=https://your-resource.services.ai.azure.com
AZURE_MISTRAL_API_KEY=...
AZURE_MISTRAL_API_VERSION=2024-05-01-preview
AZURE_MISTRAL_DEPLOYMENT_NAME=...

# Rate limits
LLM_REQUESTS_PER_MINUTE=60
EMBEDDING_BATCH_SIZE=100

Security: In production (DEBUG=False), Django will refuse to start if SECRET_KEY is set to a placeholder value.

For analysis tuning via config.yaml, see docs/configuration.md.

---

Running the Application

1. Start the Django Development Server

source .venv/bin/activate
python manage.py runserver

Open http://localhost:8000 (redirects to /dashboard/).

2. Start the Celery Worker

In a separate terminal:

source .venv/bin/activate
celery -A score worker -l info

For dev mode without Redis (using SQLAlchemy database broker):

# Make sure .env has: CELERY_BROKER_BACKEND=database
celery -A score worker -l info -P solo

The -P solo flag runs a single-threaded worker, required for the database broker.

---

Tests

source .venv/bin/activate
pytest tests/

Test configuration is in pyproject.toml:

[tool.pytest.ini_options]
DJANGO_SETTINGS_MODULE = "score.settings"

The test suite covers all major modules: analysis views, audit views, chat, chunking, claims extraction, clustering, connectors, contradictions, dashboard, duplicates, gaps, hashing, LLM client, middleware, models, pipeline integration, reports, scoring, semantic graph (NSG), tenant isolation, tracing, and vector store.

---

Dependencies

Core: Django 5.1, Celery 5.4, sqlite-vec 0.1.6, OpenAI SDK, tiktoken, django-allauth, whitenoise, gunicorn

ML/Analysis: scikit-learn, HDBSCAN, datasketch (MinHash), numpy, NLTK, langid, rank-bm25

Semantic Graph: spaCy, NetworkX, FAISS (faiss-cpu)

Document Parsing: BeautifulSoup4, pypdf, python-docx, python-pptx, markdown

PDF Export: xhtml2pdf

HTTP: httpx

Optional Connectors: msal + office365-rest-python-client (SharePoint), atlassian-python-api (Confluence)

See requirements.txt for the full list with version constraints.

---

Documentation

Detailed documentation lives in the docs/ folder:

| Document | Description | |----------|-------------| | SCORE_FORMULA.md | Scoring formula, 7 dimensions, all edge cases | | INGESTION_AND_ANALYSIS.md | Ingestion pipeline + analysis methods | | stack-and-algorithms.md | Technical stack and algorithms (French) | | deployment.md | Production deployment guide (Docker, PostgreSQL, nginx) | | project-structure.md | Full project tree with file descriptions | | configuration.md | config.yaml reference (analysis tuning) | | technical-reference.md | Celery tasks, Django apps, DB schema, URL routes, auth, scoring |

---

Contributing

See CONTRIBUTING.md for development guidelines.

---

License

This project is licensed under the Apache License 2.0 — see LICENSE for details.