ai4j

Positioning#

ai4j is a Java AI Agentic development suite for JDK 8+, providing unified LLM invocation, common AI base capabilities, and intelligent Agent development. Mainstream Java AI frameworks (Spring AI, LangChain4j) require Java 17+ and Spring Boot 3.x; ai4j fills the gap for legacy JDK 8 systems to integrate AI capabilities.

Core Capabilities#

Unified Multi-Platform LLM Access#

Aligned with OpenAI format, eliminating platform differences. Supports 9 platforms: OpenAI, Zhipu, DeepSeek, Moonshot, Hunyuan, Lingyi, Ollama (local models), MiniMax, Baichuan.

Supported service types: Chat Completions (streaming/non-streaming), Responses, Embedding, Rerank, Audio, Image, Realtime.

Full-Chain MCP Protocol#

Supports MCP Client, Server, Gateway; STDIO, SSE, Streamable HTTP transports; custom/dynamic data sources and auto-reconnection.

Tool Call (Function Calling)#

Unified interface, parallel multi-function calls, streaming function call parameter output, optimized function call loop.

RAG (Retrieval-Augmented Generation)#

• Storage: Unified VectorStore abstraction (Pinecone, Qdrant, pgvector, Milvus)
• Pipeline: IngestionPipeline (DocumentLoader → Chunker → MetadataEnricher → Embedding → VectorStore.upsert)
• Retrieval: DenseRetriever, Bm25Retriever, HybridRetriever
• Fusion: RrfFusionStrategy, RsfFusionStrategy, DbsfFusionStrategy
• Reranking: Unified IRerankService (Jina, Ollama, Doubao)
• Evaluation: Precision@K, Recall@K, F1@K, MRR, NDCG

Agent Capabilities#

• Agent Runtime (ai4j-agent): ReAct reasoning, subagent, agent teams, memory, trace, tool loop
• Coding Agent (ai4j-coding): workspace tools, outer loop, checkpoint compaction, subagent & team collaboration
• CLI / TUI / ACP (ai4j-cli): Built-in Coding Agent interface; one-shot & continuous sessions; Provider profile persistence; Session management (persist/resume/fork/history/tree/events/replay); ACP mode for IDE integration

Workflow Platform Integration#

Direct integration with Dify (Chat/Workflow), Coze (Chat/Workflow), n8n (Webhook Workflow).

Engineering Enhancements#

Unified error handling (OpenAI error type aligned), SPI mechanism (custom Dispatcher/ConnectPool), decorator enhancement (e.g., SearXNG web search), ChatMemory multi-turn context, Token statistics, multimodal (Vision), Spring Boot auto-configuration, FlowGram workflow integration.

Module Structure#

• ai4j: Core SDK (model invocation, Tool Call, MCP, RAG, VectorStore, ChatMemory)
• ai4j-agent: General Agent runtime
• ai4j-coding: Coding Agent runtime
• ai4j-cli: CLI / TUI / ACP delivery entry
• ai4j-spring-boot-starter: Spring Boot auto-configuration
• ai4j-flowgram-spring-boot-starter: FlowGram workflow integration
• ai4j-bom: Unified version management BOM

Quick Start#

Prerequisite: JDK 1.8+

CLI installation:

# Linux / macOS
curl -fsSL https://lnyo-cly.github.io/ai4j/install.sh | sh
# Windows (PowerShell)
irm https://lnyo-cly.github.io/ai4j/install.ps1 | iex

Maven dependency:

<dependencyManagement>
    <dependencies>
        <dependency>
            <groupId>io.github.lnyo-cly</groupId>
            <artifactId>ai4j-bom</artifactId>
            <version>${project.version}</version>
            <type>pom</type>
            <scope>import</scope>
        </dependency>
    </dependencies>
</dependencyManagement>
<dependency>
    <groupId>io.github.lnyo-cly</groupId>
    <artifactId>ai4j</artifactId>
</dependency>

For Spring Boot, use ai4j-spring-boot-starter.

Comparison with Alternatives#

Unconfirmed Information#

Latest release version (README uses placeholder), Hugging Face page, associated academic papers, production user cases, FlowGram independent project status.