PAI-RAG
An open-source agentic RAG framework, built on Alibaba Cloud's PAI platform, that combines knowledge-base management, multimodal document understanding, a ReAct agent and MCP tool integration for enterprise question answering.
PAI-RAG manages knowledge bases with concurrent upload, configurable chunking strategies and metadata filtering, and parses more than twenty document formats including PDF, Office files, images and video. It supports vector, full-text and hybrid retrieval with reranking, and lets a chunk be viewed, edited or toggled individually.
Its ReAct agent reasons over a think-act-observe loop with parallel tool calls, plans and breaks down multi-step tasks, and streams output over SSE. Built-in tools include knowledge-base retrieval, web search through Alibaba Cloud IQS or Tavily, page browsing, a Python code sandbox, and text-to-SQL over MySQL or PostgreSQL. MCP support lets it connect to external MCP servers or expose its own retrieval as MCP tools.
It supports multi-tenant deployments with role-based access control down to the chunk level, OpenTelemetry tracing, and an OpenAI-compatible chat completions endpoint. Supported vector backends include Chroma, Milvus and Elasticsearch.