Voxen: A Self-Hostable RAG Customer Service Bot Platform

Voxen is a self-hostable customer service bot platform based on large language models. Its core is building intelligent agents using Retrieval-Augmented Generation (RAG) technology, supporting domain-specific Q&A generation from knowledge bases, and embedding into any website via a single-line script. Key advantages include full data control (sensitive documents stored locally), a multi-module functional architecture, and support for tech stacks like Ollama (local open-source models) and Gemini (cloud-based models).

Traditional SaaS customer service solutions have the problem of lacking data control, and Voxen's self-hosted model addresses this pain point—enterprises can store sensitive documents on local infrastructure while leveraging LLM capabilities to provide intelligent Q&A services, making it especially suitable for enterprise scenarios with strict data privacy requirements.

1. Prompt Management System: Supports reusable system prompt templates to define an agent's behavior style, response format, and knowledge boundaries (e.g., technical support/sales consultation scenarios);
2. Knowledge Base and RAG Retrieval: Supports importing multiple formats such as PDF, DOCX, and web URLs, automatically chunking, vectorizing (generating 768-dimensional vectors using nomic-embed-text), and storing in the Qdrant vector database to ensure semantic retrieval accuracy;
3. Agent Construction and API Keys: Create agents by binding prompts and knowledge bases; each agent has an independent API key (vxn_...) to support multi-tenant scenarios;
4. Embedded Chat Component: A single-line script can add a floating chat button to a webpage; clicking it loads an iframe chat interface without complex integration.

Backend: Based on the FastAPI framework, uses SQLAlchemy for asynchronous operations on PostgreSQL databases, and uses Qdrant (optimized for high-dimensional vector search) for vector storage; LLM Support: Compatible with Ollama (locally deployed open-source models like Gemma3) and Google Gemini (cloud-based models); switching only requires modifying environment variables; Frontend: Built with React 19 + Vite, styled with Tailwind CSS v4 to ensure development experience and runtime performance.

Local Development: Requires Python 3.11+, PostgreSQL; if using Ollama, need to run the service locally and pull models; Docker Deployment: Provides two Compose configurations: development (docker-compose.yml with hot reloading) and production (docker-compose.prod.yml with gunicorn + nginx); Ollama service is optional (controlled via profiles); Environment Variables: All configurations are managed via .env files, including database connection URLs, LLM providers, model names, etc., facilitating cross-environment migration.

1. Small and Medium Enterprises: Controllable costs (no per-conversation SaaS fees), local data storage to avoid sensitive information leakage;
2. Developer Communities: A complete reference for RAG application implementation (document processing, vector retrieval, streaming responses, etc.);
3. Technical Teams: Plug-and-play architecture supports customization (replacing embedding models, connecting to other vector databases, adding custom authentication, etc.).

Voxen combines LLM capabilities with enterprise data sovereignty needs, making it a typical representative of self-hosted AI customer service tools. Possible future enhancement directions include multi-language support, complex conversation flow management, deep integration with existing CRM systems, etc.