@keo-ai/axiom 0.1.5 → 0.1.7

package/README.md

@@ -179,6 +179,7 @@ for await (const chunk of LLM.streamPredict({ model: 'qwen3.7-max', prompt: '讲
 - `deepseek-v4-pro`
 - `deepseek-v4-flash`
 - `kimi-k2.6`
+- `glm-5.1`
 - `qwen-vl-plus`
 
 > 目前所有模型均路由到百炼 Provider。后续接入其他厂商时，通过 `MODEL_REGISTRY` 扩展映射即可。

package/dist/function_call_loop/provider.js

@@ -47,14 +47,16 @@ function createLLMCaller(options) {
         call(messages, tools, callOptions) {
             return __awaiter(this, void 0, void 0, function* () {
                 var _a, _b;
-                const data = yield (0, llm_provider_1.callChatCompletions)(options.baseUrl, options.apiKey, {
+                const body = (0, llm_provider_1.adaptRequestForModel)({
                     model: (_a = callOptions === null || callOptions === void 0 ? void 0 : callOptions.model) !== null && _a !== void 0 ? _a : options.defaultModel,
                     messages: toOpenAIMessages(messages),
                     temperature: callOptions === null || callOptions === void 0 ? void 0 : callOptions.temperature,
                     max_tokens: callOptions === null || callOptions === void 0 ? void 0 : callOptions.maxTokens,
                     top_p: callOptions === null || callOptions === void 0 ? void 0 : callOptions.topP,
                     tools: tools.length > 0 ? toOpenAITools(tools) : undefined,
-                }, 'llm');
+                    reasoning_effort: callOptions === null || callOptions === void 0 ? void 0 : callOptions.reasoningEffort,
+                });
+                const data = yield (0, llm_provider_1.callChatCompletions)(options.baseUrl, options.apiKey, body, 'llm');
                 const choice = data.choices[0];
                 return {
                     content: choice.message.content,

package/dist/llm_provider/bailian.js

@@ -45,8 +45,8 @@ class BailianProvider {
     generate(request) {
         return __awaiter(this, void 0, void 0, function* () {
             var _a;
-            const body = this.buildRequestBody(request);
-            const data = yield (0, index_1.callChatCompletions)(this.config.baseUrl, this.config.apiKey, Object.assign(Object.assign({}, body), { stream: false }), this.name);
+            const body = (0, index_1.adaptRequestForModel)(Object.assign(Object.assign({}, this.buildRequestBody(request)), { stream: false }));
+            const data = yield (0, index_1.callChatCompletions)(this.config.baseUrl, this.config.apiKey, body, this.name);
             const choice = data.choices[0];
             return {
                 content: choice.message.content,
@@ -71,7 +71,7 @@ class BailianProvider {
         return __asyncGenerator(this, arguments, function* stream_1() {
             var _a, _b;
             const url = `${this.config.baseUrl}/chat/completions`;
-            const body = this.buildRequestBody(request);
+            const body = (0, index_1.adaptRequestForModel)(Object.assign(Object.assign({}, this.buildRequestBody(request)), { stream: true }));
             let response;
             try {
                 response = yield __await(fetch(url, {
@@ -80,7 +80,7 @@ class BailianProvider {
                         'Content-Type': 'application/json',
                         Authorization: `Bearer ${this.config.apiKey}`,
                     },
-                    body: JSON.stringify(Object.assign(Object.assign({}, body), { stream: true })),
+                    body: JSON.stringify(body),
                 }));
             }
             catch (cause) {
@@ -170,5 +170,6 @@ BailianProvider.SUPPORTED_MODELS = new Set([
     'deepseek-v4-pro',
     'deepseek-v4-flash',
     'kimi-k2.6',
+    'glm-5.1',
     'qwen-vl-plus',
 ]);

package/dist/llm_provider/index.d.ts

@@ -34,6 +34,7 @@ export interface OpenAIChatRequest {
         type: 'text' | 'json_object';
     };
     stream?: boolean;
+    reasoning_effort?: 'low' | 'medium' | 'high';
 }
 export interface OpenAIChatResponse {
     choices: Array<{
@@ -63,9 +64,14 @@ export interface OpenAIChatResponse {
  * 业务层（请求体组装、结果转换）由调用方负责。
  */
 export declare function callChatCompletions(baseUrl: string, apiKey: string, body: OpenAIChatRequest, errorPrefix?: string): Promise<OpenAIChatResponse>;
+/**
+ * 根据模型能力校验并适配 OpenAI 请求体。
+ * 若调用方明确要求了模型不支持的能力，直接抛异常。
+ */
+export declare function adaptRequestForModel(body: OpenAIChatRequest): OpenAIChatRequest;
 /** 模型枚举与注册表 */
-export type { Model, ModelConfig } from './models';
-export { MODEL_REGISTRY } from './models';
+export type { Model, ModelConfig, ModelCapabilities } from './models';
+export { MODEL_REGISTRY, MODEL_CAPABILITIES, getModelCapabilities } from './models';
 /** 标准化类型 */
 export type { Message, LLMRequest, LLMResponse, ProviderConfig, StreamChunk } from './types';
 /** Provider 接口与实现 */

package/dist/llm_provider/index.js

@@ -13,8 +13,9 @@ var __awaiter = (this && this.__awaiter) || function (thisArg, _arguments, P, ge
     });
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.BailianProvider = exports.MODEL_REGISTRY = void 0;
+exports.BailianProvider = exports.getModelCapabilities = exports.MODEL_CAPABILITIES = exports.MODEL_REGISTRY = void 0;
 exports.callChatCompletions = callChatCompletions;
+exports.adaptRequestForModel = adaptRequestForModel;
 /**
  * 发送 OpenAI 兼容的 chat completions 请求。
  * 只做 HTTP 层：构造请求、fetch、错误处理、基础解析。
@@ -56,7 +57,25 @@ function callChatCompletions(baseUrl_1, apiKey_1, body_1) {
         return data;
     });
 }
-var models_1 = require("./models");
-Object.defineProperty(exports, "MODEL_REGISTRY", { enumerable: true, get: function () { return models_1.MODEL_REGISTRY; } });
+const models_1 = require("./models");
+/**
+ * 根据模型能力校验并适配 OpenAI 请求体。
+ * 若调用方明确要求了模型不支持的能力，直接抛异常。
+ */
+function adaptRequestForModel(body) {
+    var _a;
+    const caps = (0, models_1.getModelCapabilities)(body.model);
+    if (body.reasoning_effort !== undefined && !caps.reasoningEffort) {
+        throw new Error(`[adapt] Model "${body.model}" does not support reasoning_effort`);
+    }
+    if (((_a = body.response_format) === null || _a === void 0 ? void 0 : _a.type) === 'json_object' && !caps.jsonMode) {
+        throw new Error(`[adapt] Model "${body.model}" does not support response_format json_object`);
+    }
+    return body;
+}
+var models_2 = require("./models");
+Object.defineProperty(exports, "MODEL_REGISTRY", { enumerable: true, get: function () { return models_2.MODEL_REGISTRY; } });
+Object.defineProperty(exports, "MODEL_CAPABILITIES", { enumerable: true, get: function () { return models_2.MODEL_CAPABILITIES; } });
+Object.defineProperty(exports, "getModelCapabilities", { enumerable: true, get: function () { return models_2.getModelCapabilities; } });
 var bailian_1 = require("./bailian");
 Object.defineProperty(exports, "BailianProvider", { enumerable: true, get: function () { return bailian_1.BailianProvider; } });

package/dist/llm_provider/models.d.ts

@@ -1,7 +1,7 @@
 /**
  * 支持的模型枚举。项目层通过此枚举选择模型，Axiom 内部路由到对应 Provider。
  */
-export type Model = 'qwen3.7-max' | 'qwen-plus' | 'qwen-turbo' | 'qwq-plus' | 'deepseek-v4-pro' | 'deepseek-v4-flash' | 'kimi-k2.6' | 'qwen-vl-plus';
+export type Model = 'qwen3.7-max' | 'qwen-plus' | 'qwen-turbo' | 'qwq-plus' | 'deepseek-v4-pro' | 'deepseek-v4-flash' | 'kimi-k2.6' | 'glm-5.1' | 'qwen-vl-plus';
 /** 模型到 Provider 的映射配置 */
 export interface ModelConfig {
     readonly model: string;
@@ -12,3 +12,17 @@ export interface ModelConfig {
  * 当首选 Provider 失败时，Predictor 按此表顺序尝试下一个。
  */
 export declare const MODEL_REGISTRY: Readonly<Record<Model, ReadonlyArray<ModelConfig>>>;
+/** 模型能力配置 */
+export interface ModelCapabilities {
+    /** 是否原生支持 response_format: { type: 'json_object' } */
+    readonly jsonMode: boolean;
+    /** 是否原生支持 reasoning_effort 参数 */
+    readonly reasoningEffort: boolean;
+}
+/**
+ * 模型能力矩阵。
+ * 未列出的模型默认视为不支持 jsonMode 和 reasoningEffort。
+ */
+export declare const MODEL_CAPABILITIES: Readonly<Record<string, ModelCapabilities>>;
+/** 获取指定模型的能力配置 */
+export declare function getModelCapabilities(model: string): ModelCapabilities;

package/dist/llm_provider/models.js

@@ -1,6 +1,7 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.MODEL_REGISTRY = void 0;
+exports.MODEL_CAPABILITIES = exports.MODEL_REGISTRY = void 0;
+exports.getModelCapabilities = getModelCapabilities;
 /**
  * 模型注册表。每个模型对应一个或多个 Provider 候选，按优先级排序。
  * 当首选 Provider 失败时，Predictor 按此表顺序尝试下一个。
@@ -13,5 +14,27 @@ exports.MODEL_REGISTRY = {
     'deepseek-v4-pro': [{ model: 'deepseek-v4-pro', provider: 'bailian' }],
     'deepseek-v4-flash': [{ model: 'deepseek-v4-flash', provider: 'bailian' }],
     'kimi-k2.6': [{ model: 'kimi-k2.6', provider: 'bailian' }],
+    'glm-5.1': [{ model: 'glm-5.1', provider: 'bailian' }],
     'qwen-vl-plus': [{ model: 'qwen-vl-plus', provider: 'bailian' }],
 };
+/**
+ * 模型能力矩阵。
+ * 未列出的模型默认视为不支持 jsonMode 和 reasoningEffort。
+ */
+exports.MODEL_CAPABILITIES = {
+    'qwen-max': { jsonMode: true, reasoningEffort: false },
+    'qwen3.7-max': { jsonMode: true, reasoningEffort: false },
+    'qwen-plus': { jsonMode: true, reasoningEffort: false },
+    'qwen-turbo': { jsonMode: true, reasoningEffort: false },
+    'qwq-plus': { jsonMode: true, reasoningEffort: false },
+    'deepseek-v4-pro': { jsonMode: true, reasoningEffort: false },
+    'deepseek-v4-flash': { jsonMode: true, reasoningEffort: false },
+    'kimi-k2.6': { jsonMode: true, reasoningEffort: false },
+    'glm-5.1': { jsonMode: false, reasoningEffort: false },
+    'qwen-vl-plus': { jsonMode: true, reasoningEffort: false },
+};
+/** 获取指定模型的能力配置 */
+function getModelCapabilities(model) {
+    var _a;
+    return (_a = exports.MODEL_CAPABILITIES[model]) !== null && _a !== void 0 ? _a : { jsonMode: false, reasoningEffort: false };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@keo-ai/axiom",
-  "version": "0.1.5",
+  "version": "0.1.7",
   "description": "基于 LLM 的预测与推理库，支持多 Provider 切换",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",