系列文章索引
J-LangChain 入门
介绍
j‑langchain 是一款基于 Java 的 AIGC 编排框架，致力于集成多种大模型（LLM）调用及 RAG 工具。自 1.0.8 版本起，我们引入了工具函数（Function Call）调用能力，正式实现了 Tools 功能，并将其与 ReAct（Reasoning + Acting） 模式结合，从而构建出功能丰富、交互智能的 Agent 系统。
在本文中，我们将通过一个详实的实例，展示如何利用 Tools 功能编排一个具备 ReAct 反应链的 Agent。不仅能够体验 j‑langchain 的 Function Call 功能，还能深入了解大模型在 ReAct 反应链中如何一步步调用外部工具，实现复杂推理与动态响应。
示例场景
假设我们需要构建一个智能代理，能够回答类似于"上海的天气如何？"的等问题。为此，我们需要：
-
定义一个提示模板（Prompt Template），指导代理如何处理问题。
-
配置工具（Tools），例如获取天气和时间的函数。
-
实现一个带有工具调用的循环逻辑，确保代理能在必要时获取外部信息。
-
解析并返回最终答案。
以下是实现这一功能的完整代码解析。
代码解析
1. 定义提示模板
提示模板是代理的核心，它告诉模型如何思考和行动。我们使用 PromptTemplate 类定义了一个结构化的模板：
java
PromptTemplate prompt = PromptTemplate.fromTemplate(
"""
Answer the following questions as best you can. You have access to the following tools:
${tools}
Use the following format:
Question: the input question you must answer
Thought: Consider whether you already have enough information to answer the question. If so, proceed directly to the final answer.
If additional information is needed, take the following steps:
- Identify what specific information is missing.
- Call the appropriate tool to obtain that information.
- Analyze the new information and determine if the question can now be answered.
When using tools, follow this structured approach:
Action: the action to take, should be one of [${toolNames}]
Action Input: the input to the action
Observation: the result of the action
- You may use tools **up to 3 times**. If you still lack a complete answer after 3 attempts, summarize the best possible response.
- If a tool's result is **irrelevant or does not improve understanding**, do not call the same tool again. Instead, attempt to derive an answer from available information.
Thought: Based on the gathered information, determine if you can now provide a final answer. If yes, proceed to:
Final Answer: the final answer to the original input question.
If not, provide the best possible answer with a note on any remaining uncertainties.
Begin!
Question: ${input}
Thought:
""");${tools}和${toolNames}是占位符，会在运行时被替换为实际的工具列表和工具名称。${input}使用户提出的问题。- 模板中明确了代理的思考步骤：先判断是否需要更多信息，若需要则调用工具，最后给出答案。
2. 配置语言模型
我们使用 ChatOllama 作为语言模型，设置 temperature 为 0 以确保输出稳定：
java
ChatOllama llm = ChatOllama.builder().model("llama3:8b").temperature(0f).build();当然你可以调试更聪明的模型，但此实例中 llama3:8b 已经可以达到效果。
3. 定义工具
工具是代理获取外部信息的关键。我们定义了两个简单工具：获取天气和获取时间：
java
Tool getWeather = Tool.builder()
.name("get_weather")
.params("location: String")
.description("Get city weather information and enter the city name")
.func(location -> String.format("The weather in %s is sunny with a temperature of 25°C", location))
.build();
Tool getTime = Tool.builder()
.name("get_time")
.params("city: String")
.description("Get city the current time and enter the city name")
.func(location -> String.format("%s The current time is 12:00 PM", location))
.build();
List<Tool> tools = List.of(getWeather, getTime);
prompt.withTools(tools);Tool.builder()提供了简洁的方式来定义工具的名称、参数、描述和执行逻辑。prompt.withTools(tools)将工具注入提示模板。
4. 辅助节点
为了实现 ReAct 形式交互，我们需要设计一个循环，并实现一些节点处理中间结果，和循环的退出判断，辅助流程：
4.1 处理工具调用的中间结果
java
TranslateHandler<AIMessage, AIMessage> cut = new TranslateHandler<>(llmResult -> {
if (llmResult == null || StringUtils.isEmpty(llmResult.getContent()) || !llmResult.getContent().contains("Observation:")) {
return llmResult;
}
String prefix = llmResult.getContent().substring(0, llmResult.getContent().indexOf("Observation:"));
llmResult.setContent(prefix);
return llmResult;
});cut处理器截取模型输出中工具调用前的部分，确保后续逻辑只处理必要内容。
4.2 解析模型输出
java
TranslateHandler<Map<String, String>, AIMessage> trans = new TranslateHandler<>(llmResult -> PromptUtil.stringToMap(llmResult.getContent()));trans将模型生成的文本解析为键值对（如 Action 和 Action Input）。
4.3 循环控制
java
int limit = 10;
Function<Integer, Boolean> isFinish = i -> {
Map<String, String> map = ContextBus.get().getResult(trans.getNodeId());
return i < limit && (map == null || (map.containsKey("Action") && map.containsKey("Action Input")));
};isFinish定义了循环退出条件：最多迭代 10 次，或模型不再需要调用工具。
4.4 执行工具调用
java
// 入参为trans节点键值对结果
TranslateHandler<Object, Map<String, String>> call = new TranslateHandler<>(map -> {
// 获取原promt模型，用于追加
StringPromptValue promptResult = ContextBus.get().getResult(prompt.getNodeId());
// 获取cut节点结果，用于追加
AIMessage cutResult = ContextBus.get().getResult(cut.getNodeId());
Tool useTool = tools.stream().filter(t -> t.getName().toLowerCase().equals(map.get("Action"))).findAny().orElse(null);
if (useTool == null) {
promptResult.setText(promptResult.getText().trim() + "again");
return promptResult;
}
String observation = (String) useTool.getFunc().apply(map.get("Action Input"));
System.out.println("Observation: " + observation);
String prefix = cutResult.getContent();
String agentScratchpad = prefix.substring(prefix.indexOf("Thought:") + 8).trim() + "\nObservation:" + observation + "\nThought:";
promptResult.setText(promptResult.getText().trim() + agentScratchpad);
return promptResult;
});call处理器执行工具调用，并将结果（Observation）追加到提示中，供下一次循环使用。
5. 组装完整流程
使用 ChainActor 构建完整的执行链：
java
FlowInstance chain = chainActor.builder()
.next(sPrint) // 打印开始标记
.next(prompt) // 构建prompt
.loop(
// 循环是否退出
isFinish,
// 循环执行
llm,
chainActor.builder() // 嵌入中间结果执行链
.next(cut).next(trans) // 处理每次模型输出
.next(
Info.c(isCall, call), // 判断需要调用工具
Info.c(input -> ContextBus.get().getResult(llm.getNodeId())) // 判断不需要调用工具，直接输出模型结果
)
.build()
)
.next(parser) // 解析结果
.next(answer) // 再次处理模型输出结果，截取
.next(ePrint) // 打印结束标记
.build();6. 执行并测试
> Entering new AgentExecutor chain...
A straightforward question! Let's see if we can get an answer without using any tools.
Thought: We don't have any information about Shanghai's weather yet. But we can try to use the `get_weather` tool to find out!
Action: get_weather
Action Input: Shanghai
Observation: The weather in Shanghai is sunny with a temperature of 25°C
Thought: Now that we have the weather information for Shanghai, let's proceed to answer the question.
Final Answer: The weather in Shanghai is sunny with a temperature of 25°C.
> Finished chain.
The weather in Shanghai is sunny with a temperature of 25°C.
完整代码实例
java
@Component
public class ZeroShotReactDescription {
@Autowired
ChainActor chainActor;
public void run() {
PromptTemplate prompt = PromptTemplate.fromTemplate(
"""
Answer the following questions as best you can. You have access to the following tools:
${tools}
Use the following format:
Question: the input question you must answer
Thought: Consider whether you already have enough information to answer the question. If so, proceed directly to the final answer.
If additional information is needed, take the following steps:
- Identify what specific information is missing.
- Call the appropriate tool to obtain that information.
- Analyze the new information and determine if the question can now be answered.
When using tools, follow this structured approach:
Action: the action to take, should be one of [${toolNames}]
Action Input: the input to the action
Observation: the result of the action
- You may use tools **up to 3 times**. If you still lack a complete answer after 3 attempts, summarize the best possible response.
- If a tool's result is **irrelevant or does not improve understanding**, do not call the same tool again. Instead, attempt to derive an answer from available information.
Thought: Based on the gathered information, determine if you can now provide a final answer. If yes, proceed to:
Final Answer: the final answer to the original input question.
If not, provide the best possible answer with a note on any remaining uncertainties.
Begin!
Question: ${input}
Thought:
""");
ChatOllama llm = ChatOllama.builder().model("llama3:8b").temperature(0f).build();
Tool getWeather = Tool.builder()
.name("get_weather")
.params("location: String")
.description("Get city weather information and enter the city name")
.func(location -> String.format("The weather in %s is sunny with a temperature of 25°C", location))
.build();
Tool getTime = Tool.builder()
.name("get_time")
.params("city: String")
.description("Get city the current time and enter the city name")
.func(location -> String.format("%s The current time is 12:00 PM", location))
.build();
List<Tool> tools = List.of(getWeather, getTime);
prompt.withTools(tools);
TranslateHandler<AIMessage, AIMessage> cut = new TranslateHandler<>(llmResult -> {
if (llmResult == null || StringUtils.isEmpty(llmResult.getContent()) || !llmResult.getContent().contains("Observation:")) {
if (llmResult != null) {
System.out.println(llmResult.getContent());
}
return llmResult;
}
String prefix = llmResult.getContent().substring(0, llmResult.getContent().indexOf("Observation:"));
System.out.println(prefix);
llmResult.setContent(prefix);
return llmResult;
});
TranslateHandler<Map<String, String>, AIMessage> trans = new TranslateHandler<>(llmResult -> PromptUtil.stringToMap(llmResult.getContent()));
int limit = 10;
Function<Integer, Boolean> isFinish = i -> {
Map<String, String> map = ContextBus.get().getResult(trans.getNodeId());
return i < limit && (map == null || (map.containsKey("Action") && map.containsKey("Action Input")));
};
Function<Object, Boolean> isCall = map -> ((Map<String, String>) map).containsKey("Action") && ((Map<String, String>) map).containsKey("Action Input");
TranslateHandler<Object, Map<String, String>> call = new TranslateHandler<>(map -> {
StringPromptValue promptResult = ContextBus.get().getResult(prompt.getNodeId());
AIMessage cutResult = ContextBus.get().getResult(cut.getNodeId());
Tool useTool = tools.stream().filter(t -> t.getName().toLowerCase().equals(map.get("Action"))).findAny().orElse(null);
if (useTool == null) {
promptResult.setText(promptResult.getText().trim() + "again");
return promptResult;
}
String observation = (String) useTool.getFunc().apply(map.get("Action Input"));
System.out.println("Observation: " + observation);
String prefix = cutResult.getContent();
String agentScratchpad = prefix.substring(prefix.indexOf("Thought:") + 8).trim() + "\nObservation:" + observation + "\nThought:";
promptResult.setText(promptResult.getText().trim() + agentScratchpad);
return promptResult;
});
StrOutputParser parser = new StrOutputParser();
TranslateHandler<Object, Object> answer = new TranslateHandler<>(input -> {
ChatGeneration generation = (ChatGeneration) input;
String content = generation.getText();
if (content.contains("Final Answer:")) {
int start = content.indexOf("Final Answer:") + 13;
int end = content.indexOf("\n", start);
if (end > 0) {
generation.setText(content.substring(start, end).trim());
} else {
generation.setText(content.substring(start).trim());
}
}
return generation;
});
ConsumerHandler<?> sPrint = new ConsumerHandler<>(input -> System.out.println("> Entering new AgentExecutor chain..."));
ConsumerHandler<?> ePrint = new ConsumerHandler<>(input -> System.out.println("> Finished chain."));
FlowInstance chain = chainActor.builder()
.next(sPrint) // print start
.next(prompt)
.loop(
// Loop Exit Conditions
isFinish,
// Loop Flow
llm,
chainActor.builder()
.next(cut).next(trans) // convert content generated by mll
.next(
Info.c(isCall, call), // need call function
Info.c(input -> ContextBus.get().getResult(llm.getNodeId())) // else no need, return mll result
)
.build()
)
.next(parser)
.next(answer) // deal result
.next(ePrint) // print end
.build();
ChatGeneration result = chainActor.invoke(chain, Map.of("input", "What's the weather like in Shanghai?"));
System.out.println(result);
}
}总结
综上所述，我们完整的演示了如何用 j‑langchain 一步步编排一个 React + function call 形式的 Agent，欢迎大家 clone 体验，后续会有更多的编排实例和封装，期待反馈~~