A Java agent platform that refuses to be just an API wrapper

What it does

Nexus Agent is a Java-based enterprise AI conversation platform that handles document Q&A, web search, and tool-augmented dialogue through a heavily engineered retrieval and agent pipeline. Instead of sending every question straight to a language model, it runs requests through a deterministic three-layer executor: disambiguation, knowledge-driven RAG, or an open-ended ReAct agent, each with its own retrieval strategy and guardrails. The system manages the full document lifecycle—from multi-format parsing and composable chunking to vector and keyword indexing—backed by a graph-based document structure in Neo4j.

The interesting bit

The platform treats “shadow routing” as a first-class citizen: when a user manually selects a document, the system silently runs its own knowledge router in the background and compares the results, continuously measuring routing quality without interrupting the user. It also applies an “evidence budget” that strictly limits the context window allocated to retrieved facts, and it will flat-out refuse to answer if no relevant evidence is found—an unusual level of frugality and honesty for an AI system.

Key highlights

• Three-layer executor routes queries through deterministic orchestration before deciding whether to disambiguate, retrieve knowledge, or unleash the ReAct agent.
• Dual-channel hybrid retrieval runs vector and keyword searches in parallel, fuses them with RRF, and optionally reranks before evidence is passed to the model.
• Parent-child chunk aggregation retrieves on small child blocks for recall, then promotes to parent blocks to preserve narrative context during generation.
• Composable chunking engine mixes structural, recursive, semantic, and LLM-driven splitting strategies depending on document layout.
• MCP tool protocol and declarative SKILL.md extensions let the agent discover external tools and capabilities without hard-coded function calls.

Caveats

• The operational footprint is substantial: the platform expects MySQL, PostgreSQL with PGVector, Elasticsearch, Neo4j, Redis, Kafka, MinIO, and Tika all running together.
• Documentation and UI context in the repository are primarily in Chinese, which may slow down navigation for English-only readers.
• It targets Java 25+, so teams on older LTS releases will need to upgrade or containerize carefully.

Verdict

This is for Java backend engineers who need to see how a real-world RAG and agent stack is wired—not copy-pasted from a quickstart. If you are looking for a lightweight Python microservice, this is the wrong postcode.