AI Chat Application

This repository contains my AI-powered web app. It’s still under active development, but there is already a working alpha version.

Auth Service

Responsibility: Access control and identity management.

• Authentication: JWT implementation (Access & Refresh tokens) using asymmetric encryption RS256.
• Security: Tokens are stored exclusively in HttpOnly cookies to mitigate XSS attacks.
• User Management: Handles user registration, login, and profile management in PostgreSQL.

Chat Service

Responsibility: Chat business logic and data persistence.

• Persistence: Stores chat structure, metadata, and full message history.
• Data Integrity: Uses Alembic for database schema versioning and migrations.
• Orchestration Logic: Receives client requests, validates them, and coordinates interactions with the LLM Service to generate responses.

LLM Service

Responsibility: Intelligent request orchestration, context management, and response generation.

• LLM Runtime: Direct integration with Ollama for local text generation.

• Orchestrator Agent: Acts as the intelligent core, analyzing user intent and selecting the optimal processing strategy:

  • Deep Research:
    Executes a comprehensive pipeline including data collection via the SearXNG API, dynamic RAG over retrieved web pages (parsing, chunking, reranking), and synthesis of the final response based on verified findings.

  • Knowledge Base:
    Works with local documents through a persistent RAG pipeline.

  • Direct Response:
    Generates answers directly using the model’s internal weights (Llama 3.1) when additional context is not required.

• RAG Pipeline:
  Built on Unstructured for intelligent document parsing and chunking, combined with vectorization and semantic search in Qdrant.
  To minimize hallucinations and maximize accuracy, a reranking stage filters and prioritizes only the most relevant text segments before they are fed into the LLM.

Infrastructure & Data Layer

• Gateway (NGINX): Acts as an API Gateway, routing external requests to the appropriate microservices.
• Data Storage: Hybrid persistence model — relational PostgreSQL for structured data and vector-based Qdrant for AI context storage.
• Containerization: The entire stack is deployed with Docker, ensuring service isolation and consistent environments.

Author

Created by Denys Bondarchuk. Feel free to reach out or contribute to the project!