DocMind-RAG: Full-Stack Implementation of an Enterprise-Grade RAG Intelligent Knowledge Base System

DocMind-RAG is a full-stack AI knowledge base system based on the RAG architecture, supporting multi-format document parsing, hybrid retrieval, Agent workflows, and enterprise-level multi-tenant isolation, providing a complete solution for enterprise knowledge management and large model implementation. The project uses a modern tech stack, presenting the full picture of a production-grade RAG system, which can be used as an out-of-the-box tool or a reference implementation.

In enterprise digital transformation, effective management of massive knowledge assets is a core challenge. DocMind is positioned as an out-of-the-box solution for scenarios such as enterprise knowledge management, technical document Q&A, and customer support knowledge bases, while also serving as a full-stack reference implementation for large model deployment.

Layered Architecture

Frontend layer (Vue3+TypeScript+Vite), service layer (FastAPI asynchronous backend), AI layer (DeepSeek LLM+Embedding+ReAct Agent), infrastructure layer (MySQL/Redis/Elasticsearch/Kafka/MinIO).

Hybrid Retrieval

Dual-path recall with BM25 keyword matching + vector semantic retrieval, combined with Reranker reordering to improve accuracy.

Asynchronous Processing

After document upload, it is stored via MinIO, decoupled via Kafka, parsed and chunked via LangChain, vectorized via Embedding, and finally written to Elasticsearch, supporting high concurrency.

• Multi-format support: Covers PDF/Word/Excel and other formats, parsed based on the LangChain ecosystem.
• Intelligent chunking: Combines sliding window + semantic chunking to preserve context and semantic integrity.
• Multi-turn dialogue: Automatically compresses context to retain relevant history within token budget.
• Answer traceability: Answers are accompanied by source references to ensure credibility and auditability.
• Real-time streaming output: Implemented via WebSocket/SSE for word-by-word output to enhance user experience.

ReAct Loop

Implements the Reasoning+Acting loop, supporting autonomous planning, tool calling, and observation reasoning (up to 10 rounds).

Tools and Skills

Built-in 11 tools (knowledge base search, etc.), with a registry mode for easy expansion; successful tool usage patterns are automatically saved as reusable skills.

Workflow Editor

Drag-and-drop DAG editor for visual orchestration of nodes like LLM/API/conditional judgment, enabling low-code construction of complex workflows.

• Multi-tenant isolation: Organizational-level data isolation ensures security.
• RBAC permissions: Three-level control (user→role→organization) for fine-grained access control.
• Security authentication: JWT + Redis Token blacklist, supporting logout invalidation.
• Audit and monitoring: Full operation logs; Prometheus+Grafana monitoring for key metrics.

Deployment Methods

One-click startup via Docker Compose, manual deployment, and quick startup via Windows batch processing.

Application Scenarios

Enterprise knowledge management, technical document Q&A, customer support knowledge bases, compliance review, and reference for large model implementation.

DocMind-RAG has complete functions and a modern architecture, covering core RAG capabilities and enterprise-grade features. It uses a reasonable tech stack and has high code quality (160 pytest cases). As an open-source project, it provides direct usage or reference learning value for enterprises and developers.