feat(aihr): expose model usage telemetry
This commit is contained in:
+10
-1
@@ -70,6 +70,14 @@ public final class AihrModelDto {
|
||||
public record ChatRequest(String prompt, String model, String systemPrompt) {
|
||||
}
|
||||
|
||||
public record ModelUsage(
|
||||
Integer promptTokens,
|
||||
Integer completionTokens,
|
||||
Integer totalTokens,
|
||||
long latencyMs
|
||||
) {
|
||||
}
|
||||
|
||||
public record ChatResponse(
|
||||
boolean configured,
|
||||
String providerCode,
|
||||
@@ -77,7 +85,8 @@ public final class AihrModelDto {
|
||||
String answer,
|
||||
String mode,
|
||||
String error,
|
||||
List<String> hints
|
||||
List<String> hints,
|
||||
ModelUsage usage
|
||||
) {
|
||||
}
|
||||
}
|
||||
|
||||
+30
-8
@@ -14,6 +14,7 @@ import org.dromara.aihr.domain.AihrModelDto.ConfigResponse;
|
||||
import org.dromara.aihr.domain.AihrModelDto.ProviderRequest;
|
||||
import org.dromara.aihr.domain.AihrModelDto.ProviderResponse;
|
||||
import org.dromara.aihr.domain.AihrModelDto.ProviderStatusRequest;
|
||||
import org.dromara.aihr.domain.AihrModelDto.ModelUsage;
|
||||
import org.dromara.common.satoken.utils.LoginHelper;
|
||||
import org.springframework.beans.factory.annotation.Value;
|
||||
import org.springframework.dao.DataAccessException;
|
||||
@@ -200,7 +201,8 @@ public class AihrModelSeedService {
|
||||
"AI 调用已被试点成本闸门关闭,当前返回文本兜底:先安抚业主情绪,确认事实与责任人,再承诺首次反馈时间。",
|
||||
"cost-guard-fallback",
|
||||
null,
|
||||
List.of("需要恢复真实调用时设置 AIHR_AI_RUNTIME_ENABLED=true 且 AIHR_AI_CHAT_ENABLED=true。")
|
||||
List.of("需要恢复真实调用时设置 AIHR_AI_RUNTIME_ENABLED=true 且 AIHR_AI_CHAT_ENABLED=true。"),
|
||||
null
|
||||
);
|
||||
}
|
||||
|
||||
@@ -215,13 +217,14 @@ public class AihrModelSeedService {
|
||||
List.of(
|
||||
"在 aihr_model_provider / aihr_model_config 配置 api_host、api_key、model_name 后可切换到真实 OpenAI-compatible 调用。",
|
||||
"接口不会返回 api_key。"
|
||||
)
|
||||
),
|
||||
null
|
||||
);
|
||||
}
|
||||
|
||||
try {
|
||||
String content = callOpenAiCompatible(runtime, modelName, prompt, request == null ? null : request.systemPrompt(), 0.2);
|
||||
return new ChatResponse(true, runtime.providerCode(), modelName, content, "openai-compatible", null, List.of());
|
||||
ModelCallResult call = callOpenAiCompatible(runtime, modelName, prompt, request == null ? null : request.systemPrompt(), 0.2);
|
||||
return new ChatResponse(true, runtime.providerCode(), modelName, call.content(), "openai-compatible", null, List.of(), call.usage());
|
||||
} catch (Exception e) {
|
||||
return new ChatResponse(
|
||||
true,
|
||||
@@ -230,7 +233,8 @@ public class AihrModelSeedService {
|
||||
"真实模型调用失败,已回退内置默认兜底:先确认事实、同步时限、再生成工单闭环。",
|
||||
"openai-compatible-failed",
|
||||
"模型调用失败,已回退内置兜底;请检查模型供应商配置。",
|
||||
List.of("检查 aihr_model_provider.api_host/api_key 和 aihr_model_config.model_name 是否与供应商一致。")
|
||||
List.of("检查 aihr_model_provider.api_host/api_key 和 aihr_model_config.model_name 是否与供应商一致。"),
|
||||
null
|
||||
);
|
||||
}
|
||||
}
|
||||
@@ -284,7 +288,7 @@ public class AihrModelSeedService {
|
||||
return Optional.empty();
|
||||
}
|
||||
try {
|
||||
return Optional.of(callOpenAiCompatible(runtime, runtime.modelName(), userPrompt, systemPrompt, temperature));
|
||||
return Optional.of(callOpenAiCompatible(runtime, runtime.modelName(), userPrompt, systemPrompt, temperature).content());
|
||||
} catch (Exception e) {
|
||||
log.warn("aihr llm tryChat failed, caller falls back to seed(处理错误已隐藏)");
|
||||
return Optional.empty();
|
||||
@@ -376,7 +380,8 @@ public class AihrModelSeedService {
|
||||
}
|
||||
}
|
||||
|
||||
private String callOpenAiCompatible(RuntimeConfig runtime, String modelName, String prompt, String systemPrompt, double temperature) throws Exception {
|
||||
private ModelCallResult callOpenAiCompatible(RuntimeConfig runtime, String modelName, String prompt, String systemPrompt, double temperature) throws Exception {
|
||||
long startedAt = System.nanoTime();
|
||||
ObjectNode body = objectMapper.createObjectNode();
|
||||
body.put("model", modelName);
|
||||
body.put("temperature", temperature);
|
||||
@@ -422,7 +427,21 @@ public class AihrModelSeedService {
|
||||
if (isBlank(content)) {
|
||||
throw new IllegalStateException("LLM response missing message.content");
|
||||
}
|
||||
return content;
|
||||
JsonNode usage = root.path("usage");
|
||||
ModelUsage modelUsage = new ModelUsage(
|
||||
integerUsage(usage, "prompt_tokens"),
|
||||
integerUsage(usage, "completion_tokens"),
|
||||
integerUsage(usage, "total_tokens"),
|
||||
Duration.ofNanos(System.nanoTime() - startedAt).toMillis()
|
||||
);
|
||||
return new ModelCallResult(content, modelUsage);
|
||||
}
|
||||
|
||||
private static Integer integerUsage(JsonNode usage, String field) {
|
||||
if (usage == null || usage.isMissingNode() || !usage.has(field) || !usage.get(field).canConvertToInt()) {
|
||||
return null;
|
||||
}
|
||||
return usage.get(field).intValue();
|
||||
}
|
||||
|
||||
private RuntimeConfig runtimeConfig(String requestedModel) {
|
||||
@@ -593,6 +612,9 @@ public class AihrModelSeedService {
|
||||
private record RuntimeConfig(String providerCode, String modelName, String baseUrl, String apiKey, boolean configured, String source) {
|
||||
}
|
||||
|
||||
private record ModelCallResult(String content, ModelUsage usage) {
|
||||
}
|
||||
|
||||
private record ProviderData(
|
||||
String providerName,
|
||||
String providerCode,
|
||||
|
||||
+3
@@ -48,6 +48,9 @@ class AihrSensitiveTextTest {
|
||||
String speechControllerSource = Files.readString(controllerSource("AihrSpeechController.java"));
|
||||
|
||||
assertTrue(modelSource.contains("AihrSensitiveText.forModel(prompt)"));
|
||||
assertTrue(modelSource.contains("prompt_tokens"));
|
||||
assertTrue(modelSource.contains("completion_tokens"));
|
||||
assertTrue(modelSource.contains("latencyMs"));
|
||||
assertTrue(modelSource.contains("外部响应体已隐藏"));
|
||||
assertTrue(modelSource.contains("模型调用失败,已回退内置兜底;请检查模型供应商配置。"));
|
||||
assertTrue(sopSource.contains("AihrSensitiveText.forModel(renderPrompt(prompt.template(), queryText, context))"));
|
||||
|
||||
Reference in New Issue
Block a user