RAG – Asystent Prawny

Retrieval-Augmented Generation system for Polish legal codes (Kodeksy) — criminal, civil, family, administrative, commercial, and other acts published by the Sejm. Upload PDF acts, index them into a vector database, then query with natural language questions grounded in the actual legal text.

Architecture

┌──────────┐    ┌─────────┐    ┌──────────┐    ┌────────┐
│ Frontend │───▶│   API   │───▶│  Worker  │───▶│ Qdrant │
│ (Svelte) │    │(FastAPI)│    │(consumer)│    │(vector)│
└──────────┘    └────┬────┘    └────┬─────┘    └────────┘
                     │              │
                ┌────┴────┐   ┌────┴─────┐
                │  Redis  │   │ RabbitMQ │
                │ (meta)  │   │ (queue)  │
                └─────────┘   └──────────┘
                     │
               ┌─────┴──────┐
               │  Embedder  │
               │(Qwen3-8B)  │
               └─────────────┘

Service	Port	Description
api	8000	FastAPI — query, admin, dataset endpoints
worker	—	Consumes PDF ingest jobs from RabbitMQ
evaluator	8001	Dataset generation and RAGAS evaluation
embedder	8080	Qwen3-Embedding-8B via HuggingFace TEI
qdrant	6333	Vector storage
rabbitmq	5672	Message queue for async PDF processing
redis	6379	Document metadata and chunk counts

Tech Stack

• Python 3.11, FastAPI, Uvicorn
• Qdrant — vector database
• Qwen/Qwen3-Embedding-8B — self-hosted embeddings (HuggingFace TEI)
• OpenAI API — chat completions (configurable model)
• RAGAS — evaluation (faithfulness, answer relevancy, context precision)
• RabbitMQ — async job queue (aio-pika)
• Redis — document metadata
• Svelte 5 + Vite — admin frontend
• Docker Compose — deployment

Prerequisites

• Docker & Docker Compose
• OpenAI API key
• ~16 GB RAM (embedder model)

Setup

1. Create .env in the project root:

OPENAI_API_KEY=sk-...
ADMIN_API_KEY=your-secret-admin-key

2. (Optional) Place PDF files in ./data/pdfs/ to make them available for indexing from the UI.

3. Start all services:

docker compose up -d

First startup downloads the embedding model (~2 GB) — subsequent starts use the cache in ./data/hf_cache/.

Environment Variables

Variable	Required	Default	Description
OPENAI_API_KEY	Yes	—	OpenAI API key
ADMIN_API_KEY	Yes	—	API key for admin endpoints (X-API-Key header)
OPENAI_CHAT_MODEL	No	gpt-4.1	Chat model for completions
CORS_ORIGINS	No	["http://localhost:5173"]	Allowed CORS origins (JSON list)
RABBITMQ_URL	No	amqp://rabbitmq:5672/	RabbitMQ connection URL

API Endpoints

User

Method	Path	Description
POST	/user/query	Query the RAG system with a question

Admin (requires X-API-Key header)

Method	Path	Description
POST	/admin/database/add	Upload and index a PDF
GET	/admin/database/list	List all documents
DELETE	/admin/database/remove?doc_id=	Remove a document
POST	/admin/database/reindex?doc_id=	Re-index a document
DELETE	/admin/database/clear	Clear all data

Datasets (on main API)

Method	Path	Description
GET	/dataset/list	List available PDF files
POST	/dataset/index?filename=	Index an existing PDF

Evaluator (port 8001)

Method	Path	Description
POST	/datasets/generate?doc_id=&count=	Generate QA dataset
POST	/datasets/run?dataset_name=	Run RAGAS evaluation
GET	/datasets/list	List generated datasets

Health

Method	Path	Description
GET	/health	API health check