feat(M1): single ReAct agent + streaming chat + KG form-lookup tool

- ReActAgent: AiServices streaming tool-loop (replaces 8-scene SceneSelector path) - AgentConfig: single agent = streaming Ollama(qwen2.5:14b) + KgQueryTool + per-conversation memory + L1-domain system prompt - SystemPromptService: injects L1 domain map (viw_kg_domain) into system prompt; degrades gracefully if KG views missing - KgQueryTool.findForms: read-only form-catalog lookup over local viw_ai_useful_forms (global metadata, no tenant) — interim formId resolver - AgentChatController: POST /api/agent/chat SSE stream ({type:token|reset|done|error}); reset drops pre-tool-call narration - chat.html: consume the new SSE endpoint (streamed), per-page conversationId, drop TTS-borrow path - application-saaslocal.yml: add to worktree (was master-only untracked -> profile silently hit remote DB) Verified running (saaslocal, :8099): streaming chat, tool loop returns real forms, write requests correctly report 'under development'.

feat(M1): single ReAct agent + streaming chat + KG form-lookup tool
- ReActAgent: AiServices streaming tool-loop (replaces 8-scene SceneSelector path) - AgentConfig: single agent = streaming Ollama(qwen2.5:14b) + KgQueryTool + per-conversation memory + L1-domain system prompt - SystemPromptService: injects L1 domain map (viw_kg_domain) into system prompt; degrades gracefully if KG views missing - KgQueryTool.findForms: read-only form-catalog lookup over local viw_ai_useful_forms (global metadata, no tenant) — interim formId resolver - AgentChatController: POST /api/agent/chat SSE stream ({type:token|reset|done|error}); reset drops pre-tool-call narration - chat.html: consume the new SSE endpoint (streamed), per-page conversationId, drop TTS-borrow path - application-saaslocal.yml: add to worktree (was master-only untracked -> profile silently hit remote DB) Verified running (saaslocal, :8099): streaming chat, tool loop returns real forms, write requests correctly report 'under development'.
zichun
1 parent 62db09e2
Showing 7 changed files with 402 additions and 79 deletions
src/main/java/com/xly/agent/ReActAgent.java
src/main/java/com/xly/config/AgentConfig.java
src/main/java/com/xly/service/SystemPromptService.java
src/main/java/com/xly/tool/KgQueryTool.java
src/main/java/com/xly/web/AgentChatController.java
src/main/resources/application-saaslocal.yml
src/main/resources/templates/chat.html
+package com.xly.agent;
+
+import dev.langchain4j.service.MemoryId;
+import dev.langchain4j.service.TokenStream;
+import dev.langchain4j.service.UserMessage;
+
+/**
+ * 单一 ReAct 智能体。
+ *
+ * <p>LangChain4j 的 {@code AiServices} 原生工具调用循环即 ReAct：模型自行决定是否调用工具、
+ * 调用哪个工具、以及何时给出最终答复。不再有旧的 8 场景 SceneSelector 路由——由 agent 自路由。
+ *
+ * <p>返回 {@link TokenStream} 以支持流式输出；{@link MemoryId} 绑定每个会话独立的对话记忆。
+ */
+public interface ReActAgent {
+
+    TokenStream chat(@MemoryId String conversationId, @UserMessage String userInput);
+}
+package com.xly.config;
+
+import com.xly.agent.ReActAgent;
+import com.xly.service.SystemPromptService;
+import com.xly.tool.KgQueryTool;
+import dev.langchain4j.memory.chat.MessageWindowChatMemory;
+import dev.langchain4j.model.ollama.OllamaStreamingChatModel;
+import dev.langchain4j.service.AiServices;
+import org.springframework.beans.factory.annotation.Value;
+import org.springframework.context.annotation.Bean;
+import org.springframework.context.annotation.Configuration;
+
+import java.time.Duration;
+
+/**
+ * 组装单一 ReAct agent（M1）。
+ *
+ * <p>= 流式 Ollama 模型 + 通用工具（当前只有 {@link KgQueryTool}）+ 每会话对话记忆
+ * + 注入 L1 域图的 system prompt。取代旧的「每表单一个 ToolMeta 工具 + 8 场景路由」。
+ */
+@Configuration
+public class AgentConfig {
+
+    @Value("${langchain4j.ollama.base-url}")
+    private String ollamaUrl;
+
+    @Value("${langchain4j.ollama.chat-model-name}")
+    private String chatModelName;
+
+    /** 专供 agent 的流式模型：低温度利于稳定的工具调用，较大 numPredict 避免答复被截断。 */
+    @Bean("agentStreamingModel")
+    public OllamaStreamingChatModel agentStreamingModel() {
+        return OllamaStreamingChatModel.builder()
+                .baseUrl(ollamaUrl)
+                .modelName(chatModelName)
+                .temperature(0.1)
+                .topP(0.9)
+                .numPredict(2048)
+                .timeout(Duration.ofSeconds(180))
+                .build();
+    }
+
+    @Bean
+    public ReActAgent reActAgent(SystemPromptService systemPromptService, KgQueryTool kgQueryTool) {
+        String systemPrompt = systemPromptService.buildSystemPrompt();
+        return AiServices.builder(ReActAgent.class)
+                .streamingChatModel(agentStreamingModel())
+                .tools(kgQueryTool)
+                .chatMemoryProvider(memoryId -> MessageWindowChatMemory.withMaxMessages(20))
+                .systemMessageProvider(memoryId -> systemPrompt)
+                .build();
+    }
+}
+package com.xly.service;
+
+import org.springframework.jdbc.core.JdbcTemplate;
+import org.springframework.stereotype.Service;
+
+import java.util.List;
+import java.util.Map;
+
+/**
+ * 构建单 agent 的 system prompt。
+ *
+ * <p>核心是把 L1 业务域地图（{@code viw_kg_domain}，11 个域 + 上下游流转 + 对应智能体）渲染进
+ * system prompt，作为常驻的「路由地图」——让单 agent 先判断问题属于哪个业务域、涉及哪些单据，
+ * 再决定调用哪个工具。L1 体量小且永远相关，适合常驻 prompt；L2/L3 大而稀疏，走工具按需查。
+ */
+@Service
+public class SystemPromptService {
+
+    private final JdbcTemplate jdbc;
+
+    public SystemPromptService(JdbcTemplate jdbc) {
+        this.jdbc = jdbc;
+    }
+
+    public String buildSystemPrompt() {
+        return """
+                你是「小羚羊」，小羚羊印刷 ERP 的智能助手。服务对象是印刷 / 包装行业的企业用户，\
+                帮助他们查询和（未来）操作 ERP 里的业务单据。
+
+                【业务域地图（L1 路由）】
+                下面是本 ERP 的业务域、单据规模及其上下游流转关系。回答前先据此判断用户的问题属于哪个域、\
+                可能涉及哪些单据，再决定怎么做：
+                %s
+                【可用工具】
+                - findForms(keyword)：按关键词检索业务表单目录，用来把用户口中的「单据 / 报表」定位到具体是哪一张表单\
+                （拿到 formId / moduleId）。当你不确定用户指的到底是哪张表单时，先调用它确认，不要自己编表单名。
+
+                【行为准则】
+                1. 凡是能用工具确认的事实（表单、单据、数据），一律调用工具，绝不凭空编造表单名、单据号或数据。
+                2. 始终用**简体中文**、简洁、面向业务人员回答；不要暴露内部字段名或技术细节，除非用户明确要求。
+                3. **直接给出最终答复**：不要复述你正在调用哪个工具、不要输出思考过程或任何过程性文字。
+                4. 你**当前只有只读的「表单目录检索」能力**。凡涉及新增 / 修改 / 删除 / 审核等写操作，\
+                如实告诉用户「该能力正在开发中」，绝不假装已经执行或已经生成单据。
+                5. 信息不足时主动向用户提问，不要猜测。
+                """.formatted(renderDomainMap());
+    }
+
+    private String renderDomainMap() {
+        List<Map<String, Object>> rows;
+        try {
+            rows = jdbc.queryForList(
+                    "SELECT sDomain, sAiScene, iForms, sDownstreamDomains, sUpstreamDomains " +
+                            "FROM viw_kg_domain ORDER BY iForms DESC");
+        } catch (Exception e) {
+            // KG 视图缺失时不应阻断启动——降级为空域图，agent 仍可运行。
+            org.slf4j.LoggerFactory.getLogger(SystemPromptService.class)
+                    .warn("加载 L1 业务域地图失败（viw_kg_domain 不可用），system prompt 将不含域图：{}", e.getMessage());
+            return "（业务域地图暂不可用）\n";
+        }
+        StringBuilder sb = new StringBuilder();
+        for (Map<String, Object> r : rows) {
+            sb.append("- ").append(r.get("sDomain"))
+                    .append("（").append(r.get("sAiScene")).append("，")
+                    .append(r.get("iForms")).append(" 张单据）");
+            String up = str(r.get("sUpstreamDomains"));
+            String down = str(r.get("sDownstreamDomains"));
+            if (notEmpty(up)) sb.append("  上游←[").append(up).append("]");
+            if (notEmpty(down)) sb.append("  下游→[").append(down).append("]");
+            sb.append('\n');
+        }
+        return sb.toString();
+    }
+
+    private static boolean notEmpty(String s) {
+        return s != null && !s.isBlank() && !"NULL".equalsIgnoreCase(s);
+    }
+
+    private static String str(Object o) {
+        return o == null ? "" : o.toString();
+    }
+}
+package com.xly.tool;
+
+import dev.langchain4j.agent.tool.P;
+import dev.langchain4j.agent.tool.Tool;
+import org.springframework.jdbc.core.JdbcTemplate;
+import org.springframework.stereotype.Component;
+
+import java.util.List;
+import java.util.Map;
+
+/**
+ * 知识图谱查询工具（M1）。
+ *
+ * <p>目前只暴露 {@link #findForms} 一个只读能力：按关键词检索 ERP 业务表单目录
+ * （视图 {@code viw_ai_useful_forms}，1748 张有用业务表单）。这是「表单名 → formId/moduleId」
+ * 定位的过渡实现，也是后续 KgSearch / formId 解析的雏形。
+ *
+ * <p>表单目录是**全局元数据**（对所有品牌一致），不涉及行级/租户数据，因此不需要租户过滤。
+ */
+@Component
+public class KgQueryTool {
+
+    private final JdbcTemplate jdbc;
+
+    public KgQueryTool(JdbcTemplate jdbc) {
+        this.jdbc = jdbc;
+    }
+
+    @Tool("按关键词检索 ERP 业务表单目录，返回匹配的表单名、底层数据源、所属菜单id(moduleId) 与表单id(formId)。"
+            + "当用户提到某类单据或报表、但你不确定具体是哪一张表单时，先用它来定位。最多返回 15 条。")
+    public String findForms(
+            @P("表单名或业务关键词，例如：报价 / 客户 / 库存 / 送货 / 应收 / 采购订单") String keyword) {
+
+        if (keyword == null || keyword.isBlank()) {
+            return "请提供一个表单名或业务关键词再检索。";
+        }
+        String kw = keyword.trim();
+        String like = "%" + kw + "%";
+
+        List<Map<String, Object>> rows = jdbc.queryForList(
+                "SELECT sFormTitle, sDataSource, sExecType, sModuleId, sFormId " +
+                        "FROM viw_ai_useful_forms " +
+                        "WHERE sFormTitle LIKE ? " +
+                        "ORDER BY CHAR_LENGTH(sFormTitle) ASC " +
+                        "LIMIT 15",
+                like);
+
+        if (rows.isEmpty()) {
+            return "没有找到名称包含「" + kw + "」的业务表单。可以换个更常见的说法，或告诉我更具体的单据名称。";
+        }
+
+        StringBuilder sb = new StringBuilder();
+        sb.append("找到 ").append(rows.size()).append(" 张与「").append(kw).append("」相关的业务表单：\n\n");
+        sb.append("| 表单 | 数据源(").append("表/视图/存储过程) | 菜单id | 表单id |\n");
+        sb.append("|---|---|---|---|\n");
+        for (Map<String, Object> r : rows) {
+            sb.append("| ").append(str(r.get("sFormTitle")))
+                    .append(" | ").append(str(r.get("sDataSource"))).append(" · ").append(str(r.get("sExecType")))
+                    .append(" | ").append(str(r.get("sModuleId")))
+                    .append(" | ").append(str(r.get("sFormId")))
+                    .append(" |\n");
+        }
+        return sb.toString();
+    }
+
+    private static String str(Object o) {
+        return o == null ? "" : o.toString();
+    }
+}
+package com.xly.web;
+
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.xly.agent.ReActAgent;
+import dev.langchain4j.service.TokenStream;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import org.springframework.http.MediaType;
+import org.springframework.web.bind.annotation.PostMapping;
+import org.springframework.web.bind.annotation.RequestBody;
+import org.springframework.web.bind.annotation.RequestMapping;
+import org.springframework.web.bind.annotation.RestController;
+import org.springframework.web.servlet.mvc.method.annotation.SseEmitter;
+
+import java.io.IOException;
+import java.util.LinkedHashMap;
+import java.util.Map;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+
+/**
+ * 单 agent 对话入口（M1）。
+ *
+ * <p>{@code POST /xlyAi/api/agent/chat} 以 SSE（text/event-stream）流式返回助手的 token。
+ * 每帧是一条 JSON：{@code {"type":"token|done|error","content":"..."}}，前端逐帧渲染。
+ *
+ * <p>会话记忆按 {@code conversationId} 隔离（多条命名会话）；缺省用 {@code userid:default}。
+ * 说明：token 流本身是异步的（{@link TokenStream#start()} 立即返回，回调在模型线程触发），
+ * 这里用一个线程池提交，避免占用请求线程。
+ */
+@RestController
+@RequestMapping("/api/agent")
+public class AgentChatController {
+
+    private static final Logger log = LoggerFactory.getLogger(AgentChatController.class);
+
+    private final ReActAgent agent;
+    private final ObjectMapper mapper;
+    private final ExecutorService exec = Executors.newCachedThreadPool();
+
+    public AgentChatController(ReActAgent agent, ObjectMapper mapper) {
+        this.agent = agent;
+        this.mapper = mapper;
+    }
+
+    /** 前端请求体：身份字段透传（M1 只用 userid + conversationId + text）。 */
+    public static class ChatReq {
+        public String text;
+        public String userid;
+        public String conversationId;
+    }
+
+    @PostMapping(value = "/chat", produces = "text/event-stream;charset=UTF-8")
+    public SseEmitter chat(@RequestBody ChatReq req) {
+        SseEmitter emitter = new SseEmitter(180_000L);
+        final String userInput = req.text == null ? "" : req.text;
+        final String convId = (req.conversationId != null && !req.conversationId.isBlank())
+                ? req.conversationId
+                : ((req.userid == null ? "anon" : req.userid) + ":default");
+
+        exec.submit(() -> {
+            try {
+                TokenStream ts = agent.chat(convId, userInput);
+                ts.onPartialResponse(token -> send(emitter, "token", token))
+                        // 工具执行 = 一轮结束：此前流出的是模型的“调用旁白/思考”，让前端清空，
+                        // 只保留工具执行之后的最终答复（既去掉噪声、又保留流式）。
+                        .onToolExecuted(te -> send(emitter, "reset", ""))
+                        .onCompleteResponse(resp -> {
+                            send(emitter, "done", "");
+                            emitter.complete();
+                        })
+                        .onError(err -> {
+                            log.warn("agent chat stream error (conv={})", convId, err);
+                            send(emitter, "error", err.getMessage() == null ? "服务异常" : err.getMessage());
+                            emitter.complete();
+                        })
+                        .start();
+            } catch (Exception e) {
+                log.error("agent invoke failed (conv={})", convId, e);
+                send(emitter, "error", "服务异常：" + e.getMessage());
+                emitter.complete();
+            }
+        });
+        return emitter;
+    }
+
+    private void send(SseEmitter emitter, String type, String content) {
+        try {
+            Map<String, Object> m = new LinkedHashMap<>();
+            m.put("type", type);
+            m.put("content", content);
+            emitter.send(SseEmitter.event().data(mapper.writeValueAsString(m), MediaType.APPLICATION_JSON));
+        } catch (IOException | IllegalStateException e) {
+            // 客户端已断开或 emitter 已结束——忽略。
+        }
+    }
+}
+# Local override profile so xlyAi runs against the saas-8s+ local stack on macOS.
+#
+#   Run:  JAVA_HOME=$(/usr/libexec/java_home -v 17) ./mvnw spring-boot:run \
+#             -Dspring-boot.run.profiles=saaslocal
+#
+# Only overrides what differs from the committed application.yml (which targets a
+# remote Windows deploy). Redis already points at the local shared redis-local
+# (127.0.0.1:16379 / db 0) so it is inherited unchanged. Milvus/Ollama stay remote
+# — their clients are lazy, so the app boots even if they are unreachable.
+logging:
+  dirpath: /Users/reporkey/Desktop/saas-8s+/logs/xlyAi
+
+spring:
+  datasource:
+    # Local saas DB from docker-compose.saas.yml (mysql-saas, root/local).
+    url: jdbc:mysql://127.0.0.1:33307/xlyweberp_saas?allowPublicKeyRetrieval=true&keepAlive=true&autoReconnect=true&autoReconnectForPools=true&connectTimeout=30000&socketTimeout=180000&nullCatalogMeansCurrent=true&allowMultiQueries=true&useSSL=false&useUnicode=true&characterEncoding=utf-8&failOverReadOnly=false&serverTimezone=Asia/Shanghai&zeroDateTimeBehavior=CONVERT_TO_NULL
+    username: root
+    password: local
+
+# macOS-friendly temp path (was D:/xlyweberp/ai/ocrtmp)
+ocr:
+  tmpPath: /Users/reporkey/Desktop/saas-8s+/tempPath/ocrtmp
@@ -461,6 +461,7 @@
 <script>
     let sessionId ="";
+    let conversationId = "c-" + Date.now() + "-" + Math.random().toString(36).slice(2, 8);
     let userid= "17522967560005776104370282597000";
     let username= "qianb";
     let brandsid= "1111111111";
@@ -506,35 +507,7 @@
     }
     window.onload = function(){
-        const data = {
-            text: "",
-            userid: userid,
-            username: username,
-            brandsid: brandsid,
-            subsidiaryid: subsidiaryid,
-            usertype: usertype,
-            authorization: authorization,
-            voice: "zh-CN-XiaoxiaoNeural",
-            rate: "+10%",
-            volume: "+0%",
-            voiceless: false
-        };
-
-        let initUrl=CONFIG.backendUrl+"/api/tts/init";
-        $.ajax({
-            url: initUrl,
-            type: 'POST',
-            async: false,
-            data:JSON.stringify(data),
-            dataType: 'json',
-            contentType: 'application/json; charset=UTF-8',
-            success: function(response) {
-                $("#ts").html((response.processedText + response.systemText) );
-            },
-            error: function(xhr, status, error) {
-                console.log('请求失败:', error);
-            }
-        });
+        $("#ts").html("<strong>你好，我是小羚羊 🦌</strong><br><br>我可以帮你查 ERP 里的业务单据。试试问我：<br>· “有哪些和报价相关的表单？”<br>· “客户资料在哪张表单里？”");
     }
     function reset(message){
@@ -599,70 +572,74 @@
         checkPiece();
     }
+    // ======================
+    // 单 agent 流式对话：POST /api/agent/chat -> SSE(text/event-stream)
+    // 每帧一条 JSON：{type:"token|done|error", content:"..."}
+    // ======================
     async function doMessage(input, message, button) {
         addMessage(message, 'user');
         showTypingIndicator();
-        try {
-            const requestData = {
-                text: message,
-                userid: userid,
-                usertype: usertype,
-                authorization: authorization,
-                voice: "zh-CN-XiaoxiaoNeural",
-                rate: "+10%",
-                volume: "+0%",
-                voiceless: true
-            };
+        let aiText = '';
+        let aiMsgId = null;
-            const response = await fetch(`${CONFIG.backendUrl}/api/tts/stream/query`, {
+        try {
+            const response = await fetch(`${CONFIG.backendUrl}/api/agent/chat`, {
                 method: "POST",
                 headers: { "Content-Type": "application/json;charset=UTF-8" },
-                body: JSON.stringify(requestData)
+                body: JSON.stringify({
+                    text: message,
+                    userid: userid,
+                    conversationId: conversationId
+                })
             });
-
-            const data = await response.json();
-            hideTypingIndicator();
-            const replyText = (data.processedText || "") + (data.systemText || "");
-            addMessage(replyText, 'ai');
-
-            // ==============================================
-            // 👇 【关键】用 cacheKey 取音频（绝对不串音）
-            // ==============================================
-            const cacheKey = data.cacheKey;
-            if (!cacheKey) return;
-            const audioSize = data.audioSize; // 总分几段
-
-            let retry = 0;
-            const checkAudio = async () => {
-                retry++;
-                if (retry > 20) return;
-
-                try {
-                    // ==============================================
-                    // 👇 用 cacheKey 获取自己的音频（别人拿不到）
-                    // ==============================================
-                    const res = await fetch(`${CONFIG.backendUrl}/api/tts/audio?cacheKey=${encodeURIComponent(cacheKey)}`);
-                    const audioData = await res.json();
-
-                    if (audioData.audioBase64) {
-                        const blob = base64ToBlob(audioData.audioBase64);
-                        const audio = new Audio(URL.createObjectURL(blob));
-                        audio.play().catch(err => console.log('播放异常', err));
-                    } else {
-                        setTimeout(checkAudio, 800);
+            if (!response.ok) throw new Error("HTTP " + response.status);
+
+            const reader = response.body.getReader();
+            const decoder = new TextDecoder("utf-8");
+            let buffer = '';
+
+            while (true) {
+                const { value, done } = await reader.read();
+                if (done) break;
+                buffer += decoder.decode(value, { stream: true });
+
+                let sep;
+                while ((sep = buffer.indexOf("\n\n")) >= 0) {
+                    const frame = buffer.slice(0, sep);
+                    buffer = buffer.slice(sep + 2);
+                    const dataLine = frame.split("\n").find(l => l.startsWith("data:"));
+                    if (!dataLine) continue;
+                    const payload = dataLine.slice(5).trim();
+                    if (!payload) continue;
+                    let evt;
+                    try { evt = JSON.parse(payload); } catch (e) { continue; }
+
+                    if (evt.type === "token") {
+                        if (aiMsgId === null) { hideTypingIndicator(); aiMsgId = addMessage('', 'ai'); }
+                        aiText += evt.content;
+                        updateMessage(aiMsgId, aiText);
+                    } else if (evt.type === "reset") {
+                        // 工具执行前的旁白作废，清空气泡，等最终答复流入
+                        aiText = '';
+                        if (aiMsgId === null) { hideTypingIndicator(); aiMsgId = addMessage('', 'ai'); }
+                        $(`#${aiMsgId} .message-content`).html('🔎 正在查询…');
+                    } else if (evt.type === "error") {
+                        if (aiMsgId === null) { hideTypingIndicator(); aiMsgId = addMessage('', 'ai'); }
+                        aiText += (aiText ? "\n\n" : "") + "⚠️ " + evt.content;
+                        updateMessage(aiMsgId, aiText);
                     }
-                } catch (e) {
-                    setTimeout(checkAudio, 800);
+                    // evt.type === "done" -> 结束，无需处理
                 }
-            };
-            setTimeout(checkAudio, 1200);
-            playByIndex(cacheKey, 0, audioSize);
+            }
+
+            hideTypingIndicator();
+            if (aiMsgId === null) addMessage("（无响应，请重试）", 'ai');
         } catch (error) {
             console.error('错误:', error);
             hideTypingIndicator();
-            addMessage("服务异常，请重试", 'ai');
+            if (aiMsgId === null) addMessage("服务异常，请重试：" + error.message, 'ai');
         } finally {
             input.prop('disabled', false);
             button.prop('disabled', false);
@@ -671,6 +648,11 @@
         }
     }
+    function updateMessage(messageId, content) {
+        $(`#${messageId} .message-content`).html(md.render(content));
+        scrollToBottom();
+    }
+
     // ==============================
     // 👇 语音排队播放函数（保证顺序）
     // ==============================
@@ -859,6 +841,7 @@
             localStorage.removeItem('chatHistory');
             updateStatus('对话已清空', 'connected');
             sessionId ="";
+            conversationId = "c-" + Date.now() + "-" + Math.random().toString(36).slice(2, 8);
             ensureInputAtBottom();
         }
     }