refactor(llm): switch to OpenAI-compatible API; unify model to qwen3.6-27b-iq3

- Replace langchain4j-ollama with langchain4j-open-ai; AgentFactory/ QueryTool depend on ChatModel/StreamingChatModel interfaces - OllamaJsonClient -> LlmJsonClient: raw /v1/chat/completions with response_format json_schema; single OkHttpClient instance - Thinking off at all 3 call sites via reasoning_effort=none (the only switch that works on /v1; think:false and /no_think are ignored) - Drop client-side length params (num_ctx/num_predict/max_tokens) — server-side OLLAMA_CONTEXT_LENGTH=16384 on xlyllm covers them - Unified tracing: TracingChatModelListener.record() shared by listener callbacks and LlmJsonClient; per-request model name from ctx; sql model now has the listener too (was untraced and mislabeled) - Config keys langchain4j.ollama.* -> llm.{base-url,api-key,chat-model, sql-model}; both models = qwen3.6-27b-iq3:latest (tools+thinking+ vision, verified: streaming tool-calls / json_schema / reasoning off) Verified end-to-end on :8199: intent gate traced (340/28 tokens), agent tool loop, correct answers; warm 1-4s, cold load ~96s (KEEP_ALIVE=30s).

refactor(llm): switch to OpenAI-compatible API; unify model to qwen3.6-27b-iq3
- Replace langchain4j-ollama with langchain4j-open-ai; AgentFactory/ QueryTool depend on ChatModel/StreamingChatModel interfaces - OllamaJsonClient -> LlmJsonClient: raw /v1/chat/completions with response_format json_schema; single OkHttpClient instance - Thinking off at all 3 call sites via reasoning_effort=none (the only switch that works on /v1; think:false and /no_think are ignored) - Drop client-side length params (num_ctx/num_predict/max_tokens) — server-side OLLAMA_CONTEXT_LENGTH=16384 on xlyllm covers them - Unified tracing: TracingChatModelListener.record() shared by listener callbacks and LlmJsonClient; per-request model name from ctx; sql model now has the listener too (was untraced and mislabeled) - Config keys langchain4j.ollama.* -> llm.{base-url,api-key,chat-model, sql-model}; both models = qwen3.6-27b-iq3:latest (tools+thinking+ vision, verified: streaming tool-calls / json_schema / reasoning off) Verified end-to-end on :8199: intent gate traced (340/28 tokens), agent tool loop, correct answers; warm 1-4s, cold load ~96s (KEEP_ALIVE=30s).
zichun
1 parent 4c9d29c3
Showing 11 changed files with 117 additions and 84 deletions
pom.xml
src/main/java/com/xly/config/AgentConfig.java
src/main/java/com/xly/config/AgentFactory.java
src/main/java/com/xly/config/ModelConfig.java
src/main/java/com/xly/config/TracingChatModelListener.java
src/main/java/com/xly/service/IntentService.java
src/main/java/com/xly/service/OllamaJsonClient.java → src/main/java/com/xly/service/LlmJsonClient.java
src/main/java/com/xly/tool/QueryTool.java
src/main/java/com/xly/util/OkHttpUtil.java
src/main/resources/application-saaslocal.yml
src/main/resources/application.yml
@@ -90,7 +90,7 @@
 			<version>${okhttp.version}</version>
 		</dependency>
-		<!-- LangChain4j 核心 + Ollama -->
+		<!-- LangChain4j 核心 + OpenAI 兼容协议（Ollama /v1、云端 API 皆可，走配置） -->
 		<dependency>
 			<groupId>dev.langchain4j</groupId>
 			<artifactId>langchain4j</artifactId>
@@ -98,7 +98,7 @@
 		</dependency>
 		<dependency>
 			<groupId>dev.langchain4j</groupId>
-			<artifactId>langchain4j-ollama</artifactId>
+			<artifactId>langchain4j-open-ai</artifactId>
 			<version>${langchain4j.version}</version>
 		</dependency>
 package com.xly.config;
-import dev.langchain4j.model.ollama.OllamaStreamingChatModel;
+import dev.langchain4j.model.chat.StreamingChatModel;
+import dev.langchain4j.model.openai.OpenAiStreamingChatModel;
 import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.beans.factory.annotation.Value;
 import org.springframework.context.annotation.Bean;
@@ -10,7 +11,7 @@ import java.time.Duration;
 import java.util.List;
 /**
- * agent 用的流式模型 Bean。
+ * agent 用的流式模型 Bean（OpenAI 兼容协议：Ollama /v1、vLLM、云端网关皆可，换供应商只改配置）。
  *
  * <p>单一 ReAct agent 的**组装**已移到 {@link AgentFactory}（按每次请求的身份新建携带 token/权限的工具实例，
  * 见 §5/§7 per-call context）。本类只保留全局复用的流式模型。
@@ -18,27 +19,28 @@ import java.util.List;
 @Configuration
 public class AgentConfig {
-    @Value("${langchain4j.ollama.base-url}")
-    private String ollamaUrl;
+    @Value("${llm.base-url}")
+    private String baseUrl;
-    @Value("${langchain4j.ollama.chat-model-name}")
+    @Value("${llm.api-key:ollama}")
+    private String apiKey;
+
+    @Value("${llm.chat-model}")
     private String chatModelName;
     @Autowired
     private TracingChatModelListener tracingListener;
-    /** 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。 */
+    /** 专供 agent 的流式模型：低温度利于稳定的工具调用。 */
     @Bean("agentStreamingModel")
-    public OllamaStreamingChatModel agentStreamingModel() {
-        return OllamaStreamingChatModel.builder()
-                .baseUrl(ollamaUrl)
+    public StreamingChatModel agentStreamingModel() {
+        return OpenAiStreamingChatModel.builder()
+                .baseUrl(baseUrl)
+                .apiKey(apiKey)
                 .modelName(chatModelName)
                 .temperature(0.1)
                 .topP(0.9)
-                .numPredict(2048)
-                // qwen3 支持「思考」模式，但会显著拖慢交互；关闭它 -> 快，且思考不会混进回答
-                .think(false)
-                .returnThinking(false)
+                .reasoningEffort("none")
                 .listeners(List.of(tracingListener))
                 .timeout(Duration.ofSeconds(180))
                 .build();
@@ -18,8 +18,8 @@ import com.xly.tool.ProposeWriteTool;
 import com.xly.tool.QueryTool;
 import com.xly.tool.SkillTool;
 import dev.langchain4j.memory.chat.MessageWindowChatMemory;
-import dev.langchain4j.model.ollama.OllamaChatModel;
-import dev.langchain4j.model.ollama.OllamaStreamingChatModel;
+import dev.langchain4j.model.chat.ChatModel;
+import dev.langchain4j.model.chat.StreamingChatModel;
 import dev.langchain4j.service.AiServices;
 import org.springframework.beans.factory.annotation.Qualifier;
 import org.springframework.jdbc.core.JdbcTemplate;
@@ -38,8 +38,8 @@ import org.springframework.stereotype.Component;
 @Component
 public class AgentFactory {
-    private final OllamaStreamingChatModel streamingModel;
-    private final OllamaChatModel sqlModel;
+    private final StreamingChatModel streamingModel;
+    private final ChatModel sqlModel;
     private final RedisChatMemoryStore memoryStore;
     private final SystemPromptService systemPromptService;
@@ -55,8 +55,8 @@ public class AgentFactory {
     private final SkillTool skillTool;
     private final InteractionTool interactionTool;
-    public AgentFactory(@Qualifier("agentStreamingModel") OllamaStreamingChatModel streamingModel,
-                        @Qualifier("sqlChatModel") OllamaChatModel sqlModel,
+    public AgentFactory(@Qualifier("agentStreamingModel") StreamingChatModel streamingModel,
+                        @Qualifier("sqlChatModel") ChatModel sqlModel,
                         RedisChatMemoryStore memoryStore,
                         SystemPromptService systemPromptService,
                         ErpClient erp, JdbcTemplate jdbc, FormResolverService resolver, OpService ops,
@@ -3,32 +3,39 @@ package com.xly.config;
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.fasterxml.jackson.databind.SerializationFeature;
 import com.fasterxml.jackson.datatype.jsr310.JavaTimeModule;
-import dev.langchain4j.model.ollama.OllamaChatModel;
+import dev.langchain4j.model.chat.ChatModel;
+import dev.langchain4j.model.openai.OpenAiChatModel;
 import org.springframework.beans.factory.annotation.Value;
 import org.springframework.context.annotation.Bean;
 import org.springframework.context.annotation.Configuration;
 import org.springframework.context.annotation.Primary;
 import java.time.Duration;
+import java.util.List;
 @Configuration
 public class ModelConfig {
-    @Value("${langchain4j.ollama.base-url}")
-    private String chatModelUrl;
+    @Value("${llm.base-url}")
+    private String baseUrl;
-    @Value("${langchain4j.ollama.sql-model-name}")
+    @Value("${llm.api-key:ollama}")
+    private String apiKey;
+
+    @Value("${llm.sql-model}")
     private String sqlModelName;
-    /** NL2SQL 专用模型（QueryTool 经 AgentFactory 注入）。 */
+    /** NL2SQL 专用模型（QueryTool 经 AgentFactory 注入），OpenAI 兼容协议。 */
     @Bean("sqlChatModel")
-    public OllamaChatModel sqlChatModel() {
-        return OllamaChatModel.builder()
-                .baseUrl(chatModelUrl)
+    public ChatModel sqlChatModel(TracingChatModelListener tracingListener) {
+        return OpenAiChatModel.builder()
+                .baseUrl(baseUrl)
+                .apiKey(apiKey)
                 .modelName(sqlModelName)
                 .temperature(0.0)
                 .topP(0.95)
-                .numPredict(4096)
+                .reasoningEffort("none")
+                .listeners(List.of(tracingListener))
                 .timeout(Duration.ofSeconds(120))
                 .maxRetries(3)
                 .build();
@@ -48,8 +48,6 @@ public class TracingChatModelListener implements ChatModelListener {
     private String publicKey;
     @Value("${langfuse.secret-key:}")
     private String secretKey;
-    @Value("${langchain4j.ollama.chat-model-name:unknown}")
-    private String modelName;
     public TracingChatModelListener(ObjectMapper mapper) {
         this.mapper = mapper;
@@ -59,11 +57,12 @@ public class TracingChatModelListener implements ChatModelListener {
     public void onRequest(ChatModelRequestContext ctx) {
         ctx.attributes().put("t0", System.nanoTime());
         ctx.attributes().put("startTs", Instant.now().toString());
+        String model = ctx.chatRequest() == null ? null : ctx.chatRequest().modelName();
+        ctx.attributes().put("model", model == null ? "unknown" : model);
     }
     @Override
     public void onResponse(ChatModelResponseContext ctx) {
-        long ms = elapsedMs(ctx.attributes().get("t0"));
         Integer in = null;
         Integer out = null;
         try {
@@ -74,20 +73,31 @@ public class TracingChatModelListener implements ChatModelListener {
             }
         } catch (Exception ignore) {
         }
-        log.info("LLM ok {}ms tokens(in/out)={}/{}", ms, in, out);
-        exportToLangfuse(String.valueOf(ctx.attributes().get("startTs")), in, out, null);
+        record(String.valueOf(ctx.attributes().get("model")), ctx.attributes().get("t0"),
+                String.valueOf(ctx.attributes().get("startTs")), in, out, null);
     }
     @Override
     public void onError(ChatModelErrorContext ctx) {
         Throwable e = ctx.error();
         String msg = e == null ? "?" : e.getMessage();
-        log.warn("LLM error: {}", msg);
-        exportToLangfuse(String.valueOf(ctx.attributes().get("startTs")), null, null, msg);
+        record(String.valueOf(ctx.attributes().get("model")), ctx.attributes().get("t0"),
+                String.valueOf(ctx.attributes().get("startTs")), null, null, msg);
+    }
+
+    /** 统一记录一次 LLM 调用（listener 回调与 {@code LlmJsonClient} 直连路径共用）。 */
+    public void record(String model, Object startNanos, String startTs, Integer in, Integer out, String error) {
+        long ms = elapsedMs(startNanos);
+        if (error == null) {
+            log.info("LLM ok {} {}ms tokens(in/out)={}/{}", model, ms, in, out);
+        } else {
+            log.warn("LLM error {} {}ms: {}", model, ms, error);
+        }
+        exportToLangfuse(model, startTs, in, out, error);
     }
     /** 把一条 generation span 转发到 Langfuse（best-effort，异步，失败仅告警）。未启用则直接返回。 */
-    private void exportToLangfuse(String startTs, Integer in, Integer out, String error) {
+    private void exportToLangfuse(String modelName, String startTs, Integer in, Integer out, String error) {
         if (!langfuseEnabled || publicKey == null || publicKey.isBlank() || secretKey == null || secretKey.isBlank()) {
             return;
         }
@@ -13,7 +13,7 @@ import java.util.Map;
 /**
  * 第 0 阶段意图门（intent gate）：把一句用户话分类为 {意图, 单据类型, 带角色的实体, 缺失信息}。
  *
- * <p>只做**一件窄任务**，用受约束 JSON 解码（{@link OllamaJsonClient}）。实测在此形态下
+ * <p>只做**一件窄任务**，用受约束 JSON 解码（{@link LlmJsonClient}）。实测在此形态下
  * qwen3:14b 对包括「报价纸盒→纸盒是产品而非客户」「给苏州华为报价彩盒→客户+产品分离」
  * 「有多少个客户→查询」在内的样本 8/8 正确、~2-3s/次。相较之下，让同一个模型在全部 9 个工具的
  * 单次 ReAct 里同时判意图/选工具/编参数则频繁出错（把查询错当新增、更新流程死循环等）。
@@ -39,9 +39,9 @@ public class IntentService {
             + "⑤missing 里列出完成该意图还缺的关键信息(如 修改缺『具体记录』『新值』)。"
             + "只输出 JSON，不要解释。";
-    private final OllamaJsonClient llm;
+    private final LlmJsonClient llm;
-    public IntentService(OllamaJsonClient llm) {
+    public IntentService(LlmJsonClient llm) {
         this.llm = llm;
     }
@@ -2,78 +2,93 @@ package com.xly.service;
 import com.fasterxml.jackson.databind.JsonNode;
 import com.fasterxml.jackson.databind.ObjectMapper;
+import com.xly.config.TracingChatModelListener;
 import com.xly.util.OkHttpUtil;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 import org.springframework.beans.factory.annotation.Value;
 import org.springframework.stereotype.Service;
+import java.time.Instant;
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
 /**
- * 直连 Ollama /api/chat 的**受约束 JSON**补全（constrained decoding）。
+ * OpenAI 兼容协议（/v1/chat/completions）的**受约束 JSON**补全（constrained decoding）。
  *
- * <p>为什么绕开 LangChain4j：Ollama 的 {@code format=<JSON Schema>} 会用 XGrammar 做语法约束解码，
- * 保证输出**一定**是 schema 合法的 JSON（违反 schema 的 token 概率直接置 0）——这是根治「把产品名塞进客户字段」
- * 这类**槽位/参数幻觉**的关键手段。实测 qwen3:14b 在此模式下意图/实体抽取 8/8 正确、~2-3s/次（think=false）。
+ * <p>用 {@code response_format={type:"json_schema", json_schema:{schema}}} 做语法约束解码，
+ * 保证输出**一定**是 schema 合法的 JSON（Ollama 侧由 XGrammar 实现，违反 schema 的 token 概率直接置 0）——
+ * 这是根治「把产品名塞进客户字段」这类**槽位/参数幻觉**的关键手段。实测 qwen3:14b 在此模式下
+ * 意图/实体抽取 8/8 正确、~2-3s/次（reasoning_effort=none）。
  *
- * <p>用于两处「窄而稳」的推理子任务：{@link IntentService}（意图+实体分类）与
- * {@link SlotFillService}（按表单真实字段做受约束的槽位填充）。主对话/查询仍走 LangChain4j 工具循环。
+ * <p>用于两处「窄而稳」的推理子任务：{@link IntentService}（意图+实体分类）与受约束槽位填充。
+ * 主对话/查询走 LangChain4j 工具循环。
  */
 @Service
-public class OllamaJsonClient {
+public class LlmJsonClient {
-    private static final Logger log = LoggerFactory.getLogger(OllamaJsonClient.class);
+    private static final Logger log = LoggerFactory.getLogger(LlmJsonClient.class);
     private final ObjectMapper mapper;
+    private final TracingChatModelListener tracing;
+    private final OkHttpUtil http = OkHttpUtil.getInstance(10, 120, 30);
-    @Value("${langchain4j.ollama.base-url}")
+    @Value("${llm.base-url}")
     private String baseUrl;
-    @Value("${langchain4j.ollama.chat-model-name}")
+    @Value("${llm.api-key:ollama}")
+    private String apiKey;
+
+    @Value("${llm.chat-model}")
     private String model;
-    public OllamaJsonClient(ObjectMapper mapper) {
+    public LlmJsonClient(ObjectMapper mapper, TracingChatModelListener tracing) {
         this.mapper = mapper;
+        this.tracing = tracing;
     }
     /**
-     * 受约束 JSON 补全：think=false（快且稳定，实测无精度损失）、低温度、format=schema、非流式。
+     * 受约束 JSON 补全：低温度、response_format=json_schema、非流式。
      *
      * @param system 系统提示（角色 + 抽取规则）
      * @param user   用户话
-     * @param schema JSON Schema（Map 结构，直接序列化进 format）
+     * @param schema JSON Schema（Map 结构，序列化进 response_format.json_schema.schema）
      * @return 解析后的 JsonNode；失败返回 null（调用方须降级处理，绝不因它中断主流程）
      */
     public JsonNode completeJson(String system, String user, Map<String, Object> schema) {
+        long t0 = System.nanoTime();
+        String startTs = Instant.now().toString();
         try {
             Map<String, Object> body = new LinkedHashMap<>();
             body.put("model", model);
             body.put("stream", false);
-            body.put("think", false);
-            body.put("format", schema);
+            body.put("temperature", 0.1);
+            body.put("top_p", 0.9);
+            body.put("reasoning_effort", "none");
+            body.put("response_format", Map.of(
+                    "type", "json_schema",
+                    "json_schema", Map.of("name", "output", "schema", schema)));
             body.put("messages", List.of(
                     Map.of("role", "system", "content", system),
                     Map.of("role", "user", "content", user)));
-            Map<String, Object> opts = new LinkedHashMap<>();
-            opts.put("temperature", 0.1);
-            opts.put("top_p", 0.9);
-            opts.put("num_ctx", 8192); // 足够容纳字段清单/历史，避免长上下文时工具/JSON 退化
-            body.put("options", opts);
             String json = mapper.writeValueAsString(body);
-            String resp = OkHttpUtil.getInstance(10, 120, 30).postJson(baseUrl + "/api/chat", json);
+            String resp = http.postJson(baseUrl + "/chat/completions", apiKey, json);
             JsonNode root = mapper.readTree(resp);
-            String content = root.path("message").path("content").asText("");
+            JsonNode usage = root.path("usage");
+            tracing.record(model, t0, startTs,
+                    usage.path("prompt_tokens").isNumber() ? usage.path("prompt_tokens").asInt() : null,
+                    usage.path("completion_tokens").isNumber() ? usage.path("completion_tokens").asInt() : null,
+                    null);
+            String content = root.path("choices").path(0).path("message").path("content").asText("");
             if (content.isBlank()) {
-                log.warn("ollama json completion: empty content");
+                log.warn("llm json completion: empty content");
                 return null;
             }
             return mapper.readTree(content);
         } catch (Exception e) {
-            log.warn("ollama json completion failed: {}", e.getMessage());
+            tracing.record(model, t0, startTs, null, null, e.getMessage());
             return null;
         }
     }
@@ -5,7 +5,7 @@ import com.xly.service.AuditService;
 import com.xly.service.FormResolverService;
 import dev.langchain4j.agent.tool.P;
 import dev.langchain4j.agent.tool.Tool;
-import dev.langchain4j.model.ollama.OllamaChatModel;
+import dev.langchain4j.model.chat.ChatModel;
 import net.sf.jsqlparser.expression.StringValue;
 import net.sf.jsqlparser.expression.operators.conditional.AndExpression;
 import net.sf.jsqlparser.expression.operators.relational.EqualsTo;
@@ -41,13 +41,13 @@ import java.util.regex.Pattern;
  */
 public class QueryTool {
-    private final OllamaChatModel sqlModel;
+    private final ChatModel sqlModel;
     private final JdbcTemplate jdbc;
     private final AuditService audit;
     private final AgentIdentity identity;
     private final FormResolverService resolver;
-    public QueryTool(OllamaChatModel sqlModel, JdbcTemplate jdbc, AuditService audit,
+    public QueryTool(ChatModel sqlModel, JdbcTemplate jdbc, AuditService audit,
                      AgentIdentity identity, FormResolverService resolver) {
         this.sqlModel = sqlModel;
         this.jdbc = jdbc;
@@ -6,7 +6,7 @@ import java.io.IOException;
 import java.util.concurrent.TimeUnit;
 /**
- * OkHttp 薄封装：目前仅 {@code OllamaJsonClient} 用它同步 POST JSON。
+ * OkHttp 薄封装：目前仅 {@code LlmJsonClient} 用它同步 POST JSON。
  */
 public class OkHttpUtil {
@@ -24,9 +24,14 @@ public class OkHttpUtil {
         return new OkHttpUtil(connectTimeout, readTimeout, writeTimeout);
     }
-    public String postJson(String url, String json) throws IOException {
+    /** POST JSON，带 Bearer 鉴权（OpenAI 兼容端点；Ollama 忽略该头，云端网关需要真实 key）。 */
+    public String postJson(String url, String bearerToken, String json) throws IOException {
         RequestBody body = RequestBody.create(json, MediaType.parse("application/json; charset=utf-8"));
-        Request request = new Request.Builder().url(url).post(body).build();
+        Request.Builder reqBuilder = new Request.Builder().url(url).post(body);
+        if (bearerToken != null && !bearerToken.isBlank()) {
+            reqBuilder.header("Authorization", "Bearer " + bearerToken);
+        }
+        Request request = reqBuilder.build();
         try (Response response = client.newCall(request).execute()) {
             if (!response.isSuccessful()) {
                 throw new IOException("Unexpected code: " + response.code() + ", message: " + response.message());
@@ -21,12 +21,6 @@ spring:
 # same local DB. The committed application.yml erp.baseurl targets the remote deploy.
 # dev-login lets xlyAi mint a working ERP session locally (admin/666666, brand 1111111111);
 # production should instead pass through the user's own browser ERP token per request.
-langchain4j:
-  ollama:
-    # qwen3:14b follows instructions better and keeps its chain-of-thought on a
-    # separate channel (Ollama "thinking"), so it doesn't leak into the streamed content.
-    chat-model-name: qwen3:14b
-
 erp:
   baseurl: http://127.0.0.1:8697/xlyEntry
   dev-login:
@@ -61,13 +61,13 @@ management:
     health:
       show-details: always
-langchain4j:
-  ollama:
-    base-url: http://112.82.245.194:41434
-    # 聊天模型（saaslocal 覆盖为 qwen3:14b）
-    chat-model-name: qwen2.5:14b
-    # SQL/代码模型（QueryTool NL2SQL 专用）
-    sql-model-name: qwen2.5-coder:7b
+# LLM：OpenAI 兼容协议（Ollama /v1、vLLM、云端网关皆可，换供应商只改这里）
+llm:
+  base-url: http://112.82.245.194:41434/v1
+  api-key: ollama          # Ollama 不校验（任意非空）；云端填真实 key
+  chat-model: qwen3.6-27b-iq3:latest
+  # SQL/代码模型（QueryTool NL2SQL 专用）
+  sql-model: qwen3.6-27b-iq3:latest
 erp:
   baseurl: http://118.178.19.35:8080/xlyEntry_saas