@sogni-ai/sogni-client 4.2.0-alpha.2 → 4.2.0-alpha.21

package/src/Chat/index.ts

@@ -14,6 +14,8 @@ import {
   ChatJobStateEvent,
   ChatRequestMessage,
   ChatMessage,
+  HostedChatCompletionParams,
+  HostedChatCompletionResult,
   LLMCostEstimation,
   LLMEstimateResponse,
   LLMModelInfo,
@@ -22,6 +24,52 @@ import {
 } from './types';
 import getUUID from '../lib/getUUID';
 import type ProjectsApi from '../Projects';
+import { parseInlineMediaDataUri } from '../lib/mediaValidation';
+
+const MAX_VISION_IMAGE_COUNT = 20;
+const MAX_VISION_IMAGE_BYTES = 10 * 1024 * 1024;
+const MAX_VISION_IMAGE_LONGEST_SIDE = 1024;
+
+function normalizeVisionImageDataUri(input: string): string {
+  parseInlineMediaDataUri(input, 'image', {
+    maxBytes: MAX_VISION_IMAGE_BYTES,
+    maxImageLongestSide: MAX_VISION_IMAGE_LONGEST_SIDE
+  });
+  return input.trim();
+}
+
+function normalizeVisionMessages(messages: ChatMessage[]): ChatMessage[] {
+  let imageCount = 0;
+  return messages.map((msg) => {
+    if (!Array.isArray(msg.content)) {
+      return msg;
+    }
+
+    return {
+      ...msg,
+      content: msg.content.map((part) => {
+        if (part.type !== 'image_url') {
+          return part;
+        }
+
+        imageCount += 1;
+        if (imageCount > MAX_VISION_IMAGE_COUNT) {
+          throw new Error(
+            `A maximum of ${MAX_VISION_IMAGE_COUNT} vision images is allowed per request`
+          );
+        }
+
+        return {
+          type: 'image_url' as const,
+          image_url: {
+            url: normalizeVisionImageDataUri(part.image_url.url),
+            ...(part.image_url.detail && { detail: part.image_url.detail })
+          }
+        };
+      })
+    };
+  });
+}
 
 export interface ChatApiEvents {
   /** Emitted for each token chunk received during streaming */
@@ -88,6 +136,10 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
       ((params: ChatCompletionParams) => Promise<ChatStream | ChatCompletionResult>);
   };
 
+  hosted: {
+    create: (params: HostedChatCompletionParams) => Promise<HostedChatCompletionResult>;
+  };
+
   constructor(config: ApiConfig, projects?: ProjectsApi) {
     super(config);
 
@@ -102,6 +154,9 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
     this.completions = {
       create: this.createCompletion.bind(this) as any
     };
+    this.hosted = {
+      create: this.createHostedCompletion.bind(this)
+    };
 
     // Set up the tools API (requires ProjectsApi for media generation).
     // When ProjectsApi is not provided, tool execution methods will throw at runtime.
@@ -156,7 +211,8 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
    * const estimate = await sogni.chat.estimateCost({
    *   model: 'qwen3.6-35b-a3b-gguf-iq4xs',
    *   messages: [{ role: 'user', content: 'Hello!' }],
-   *   max_tokens: 1024,
+   *   think: true,
+   *   taskProfile: 'reasoning',
    * });
    * console.log(`Estimated cost: ${estimate.costInToken.toFixed(6)}`);
    * ```
@@ -166,12 +222,15 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
     messages: ChatMessage[];
     max_tokens?: number;
     tokenType?: 'sogni' | 'spark';
+    think?: boolean;
+    taskProfile?: 'general' | 'coding' | 'reasoning';
   }): Promise<LLMCostEstimation> {
+    const normalizedMessages = normalizeVisionMessages(params.messages);
     const tokenType = params.tokenType || 'sogni';
     const inputTokens = Math.ceil(
-      JSON.stringify(this.stripImageDataForEstimation(params.messages)).length / 4
+      JSON.stringify(this.stripImageDataForEstimation(normalizedMessages)).length / 4
     );
-    const maxOutputTokens = params.max_tokens || 4096;
+    const maxOutputTokens = this.resolveEstimatedMaxOutputTokens(params);
     const pathParams = [tokenType, params.model, inputTokens, maxOutputTokens];
     const path = pathParams.map((p) => encodeURIComponent(p)).join('/');
     const r = await this.client.socket.get<LLMEstimateResponse>(`/api/v1/job-llm/estimate/${path}`);
@@ -185,6 +244,38 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
     };
   }
 
+  private resolveEstimatedMaxOutputTokens(params: {
+    model: string;
+    max_tokens?: number;
+    think?: boolean;
+    taskProfile?: 'general' | 'coding' | 'reasoning';
+  }): number {
+    if (typeof params.max_tokens === 'number' && Number.isFinite(params.max_tokens)) {
+      return params.max_tokens;
+    }
+
+    const modelInfo = this.availableLLMModels[params.model];
+    const defaultFromModel = modelInfo?.maxOutputTokens?.default;
+    const thinkingComplexDefault = modelInfo?.maxOutputTokens?.thinkingComplexDefault;
+    const isComplexThinking =
+      params.think === true &&
+      (params.taskProfile === 'coding' || params.taskProfile === 'reasoning');
+
+    if (
+      isComplexThinking &&
+      typeof thinkingComplexDefault === 'number' &&
+      Number.isFinite(thinkingComplexDefault)
+    ) {
+      return thinkingComplexDefault;
+    }
+
+    if (typeof defaultFromModel === 'number' && Number.isFinite(defaultFromModel)) {
+      return defaultFromModel;
+    }
+
+    return 4096;
+  }
+
   private handleSwarmLLMModels(data: Record<string, number | LLMModelInfo>): void {
     const models: Record<string, LLMModelInfo> = {};
     for (const [modelId, value] of Object.entries(data)) {
@@ -233,6 +324,14 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
     return { enable_thinking: think };
   }
 
+  private normalizeSogniToolsMode(
+    value: ChatCompletionParams['sogni_tools'] | string | undefined
+  ): ChatRequestMessage['sogni_tools'] | undefined {
+    return typeof value === 'string' && value.trim().toLowerCase() === 'rich'
+      ? 'creative-tools'
+      : (value as ChatRequestMessage['sogni_tools']);
+  }
+
   private async createCompletion(
     params: ChatCompletionParams
   ): Promise<ChatStream | ChatCompletionResult> {
@@ -250,6 +349,46 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
     return this.createSingleCompletion(params);
   }
 
+  private async createHostedCompletion(
+    params: HostedChatCompletionParams
+  ): Promise<HostedChatCompletionResult> {
+    if (params.stream) {
+      throw new Error('chat.hosted.create currently supports non-streaming requests only.');
+    }
+
+    const normalizedMessages = normalizeVisionMessages(params.messages);
+    const chatTemplateKwargs =
+      params.chat_template_kwargs ?? this.buildChatTemplateKwargs(params.think);
+    return this.client.rest.post<HostedChatCompletionResult>(
+      '/v1/chat/completions',
+      {
+        model: params.model,
+        messages: normalizedMessages,
+        app_source: params.app_source || params.appSource || this.client.appSource,
+        max_tokens: params.max_tokens,
+        temperature: params.temperature,
+        top_p: params.top_p,
+        top_k: params.top_k,
+        min_p: params.min_p,
+        repetition_penalty: params.repetition_penalty,
+        frequency_penalty: params.frequency_penalty,
+        presence_penalty: params.presence_penalty,
+        stop: params.stop,
+        token_type: params.token_type || params.tokenType,
+        tools: params.tools,
+        tool_choice: params.tool_choice,
+        sogni_tools: this.normalizeSogniToolsMode(params.sogni_tools),
+        sogni_tool_execution: params.sogni_tool_execution,
+        task_profile: params.taskProfile,
+        media_references: params.media_references ?? params.mediaReferences,
+        api_media_references: params.api_media_references ?? params.apiMediaReferences,
+        ...(chatTemplateKwargs && { chat_template_kwargs: chatTemplateKwargs }),
+        ...(params.response_format && { response_format: params.response_format })
+      },
+      { timeoutMs: 300000 }
+    );
+  }
+
   /**
    * Send a single chat completion request (no auto tool execution).
    */
@@ -257,6 +396,7 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
     params: ChatCompletionParams
   ): Promise<ChatStream | ChatCompletionResult> {
     const jobID = getUUID();
+    const normalizedMessages = normalizeVisionMessages(params.messages);
 
     // Build chat_template_kwargs from think parameter
     const chatTemplateKwargs = this.buildChatTemplateKwargs(params.think);
@@ -265,7 +405,8 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
       jobID,
       type: 'llm',
       model: params.model,
-      messages: params.messages,
+      messages: normalizedMessages,
+      appSource: params.appSource || this.client.appSource,
       max_tokens: params.max_tokens,
       temperature: params.temperature,
       top_p: params.top_p,
@@ -279,8 +420,11 @@ class ChatApi extends ApiGroup<ChatApiEvents> {
       tokenType: params.tokenType,
       tools: params.tools,
       tool_choice: params.tool_choice,
+      sogni_tools: this.normalizeSogniToolsMode(params.sogni_tools),
+      sogni_tool_execution: params.sogni_tool_execution,
       taskProfile: params.taskProfile,
-      ...(chatTemplateKwargs && { chat_template_kwargs: chatTemplateKwargs })
+      ...(chatTemplateKwargs && { chat_template_kwargs: chatTemplateKwargs }),
+      ...(params.response_format && { response_format: params.response_format })
     };
 
     const stream = new ChatStream(jobID);