Ragent AI

Ragent AI is an open-source Java AI application platform for Agentic RAG, published by the nageoffer organization and documented by its author, 程序员马丁. It covers the full path from document ingestion to question answering, with a React console for end users and administrators. The latest GitHub release listed is 1.1.0, and the code is licensed under Apache-2.0.

What It Is

Ragent is a modular-monolith application with a Spring Boot backend and a React 18 frontend. Its backend is split into seven Maven modules: model infrastructure, system, RAG, agent, bootstrap, framework and a sample MCP server. The project positions itself as a source-level engineering reference for building Java AI applications.

Retrieval and Question Understanding

Vector, keyword (Elasticsearch), knowledge-graph (LightRAG with Neo4j) and web-search channels can run in parallel, depending on configuration. Results go through deduplication, weighted RRF fusion, Rerank and metadata enrichment. Before retrieval, the system supports query-term mapping, question rewriting and splitting, tree-structured intent recognition and routing across multiple knowledge bases. Documents enter through a configurable ingestion pipeline of fetch, parse, enhance, chunk, enrich and index nodes, with per-node execution logs.

Workflow and Agent Modes

The docs describe two coexisting paths. V1 is a fixed Workflow: rewrite, intent recognition, multi-channel recall, then generation. V2 is an agentic path built on AgentScope's ReAct loop, where knowledge retrieval, MCP tools and Skills are tools the model chooses from. Write operations can require human confirmation, and agent context is trimmed and compressed with long-term memory across sessions.

Reliability and Operations

Model routing offers model tiers, first-packet probing and a three-state circuit breaker with automatic fallback. Redis-based fair queuing protects model services from bursts. Answers keep source citations, and the platform provides user feedback, traces, audit logs with config diffs, and Langfuse integration for tracing reasoning, tool calls and token usage. Model providers named include Bailian, SiliconFlow, AIHubMix and Ollama. PostgreSQL with pgvector is the default vector store, with Milvus as an alternative.