Merge pull request #6292 from Little-LittleProgrammer/feature/alibaba-omni-support

feat(alibaba): Added alibaba vision model and omni model support

Author: Leizhenpeng

app/client/api.ts

@@ -40,6 +40,11 @@ export interface MultimodalContent {
   };
 }
 
+export interface MultimodalContentForAlibaba {
+  text?: string;
+  image?: string;
+}
+
 export interface RequestMessage {
   role: MessageRole;
   content: string | MultimodalContent[];

app/client/platforms/alibaba.ts

@@ -7,20 +7,25 @@ import {
   ChatMessageTool,
   usePluginStore,
 } from "@/app/store";
-import { streamWithThink } from "@/app/utils/chat";
+import {
+  preProcessImageContentForAlibabaDashScope,
+  streamWithThink,
+} from "@/app/utils/chat";
 import {
   ChatOptions,
   getHeaders,
   LLMApi,
   LLMModel,
   SpeechOptions,
   MultimodalContent,
+  MultimodalContentForAlibaba,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import {
   getMessageTextContent,
   getMessageTextContentWithoutThinking,
   getTimeoutMSByModel,
+  isVisionModel,
 } from "@/app/utils";
 import { fetch } from "@/app/utils/stream";
 
@@ -89,14 +94,6 @@ export class QwenApi implements LLMApi {
   }
 
   async chat(options: ChatOptions) {
-    const messages = options.messages.map((v) => ({
-      role: v.role,
-      content:
-        v.role === "assistant"
-          ? getMessageTextContentWithoutThinking(v)
-          : getMessageTextContent(v),
-    }));
-
     const modelConfig = {
       ...useAppConfig.getState().modelConfig,
       ...useChatStore.getState().currentSession().mask.modelConfig,
@@ -105,6 +102,21 @@ export class QwenApi implements LLMApi {
       },
     };
 
+    const visionModel = isVisionModel(options.config.model);
+
+    const messages: ChatOptions["messages"] = [];
+    for (const v of options.messages) {
+      const content = (
+        visionModel
+          ? await preProcessImageContentForAlibabaDashScope(v.content)
+          : v.role === "assistant"
+          ? getMessageTextContentWithoutThinking(v)
+          : getMessageTextContent(v)
+      ) as any;
+
+      messages.push({ role: v.role, content });
+    }
+
     const shouldStream = !!options.config.stream;
     const requestPayload: RequestPayload = {
       model: modelConfig.model,
@@ -129,7 +141,7 @@ export class QwenApi implements LLMApi {
         "X-DashScope-SSE": shouldStream ? "enable" : "disable",
       };
 
-      const chatPath = this.path(Alibaba.ChatPath);
+      const chatPath = this.path(Alibaba.ChatPath(modelConfig.model));
       const chatPayload = {
         method: "POST",
         body: JSON.stringify(requestPayload),
@@ -162,7 +174,7 @@ export class QwenApi implements LLMApi {
             const json = JSON.parse(text);
             const choices = json.output.choices as Array<{
               message: {
-                content: string | null;
+                content: string | null | MultimodalContentForAlibaba[];
                 tool_calls: ChatMessageTool[];
                 reasoning_content: string | null;
               };
@@ -212,7 +224,9 @@ export class QwenApi implements LLMApi {
             } else if (content && content.length > 0) {
               return {
                 isThinking: false,
-                content: content,
+                content: Array.isArray(content)
+                  ? content.map((item) => item.text).join(",")
+                  : content,
               };
             }
 

app/constant.ts

@@ -221,7 +221,12 @@ export const ByteDance = {
 
 export const Alibaba = {
   ExampleEndpoint: ALIBABA_BASE_URL,
-  ChatPath: "v1/services/aigc/text-generation/generation",
+  ChatPath: (modelName: string) => {
+    if (modelName.includes("vl") || modelName.includes("omni")) {
+      return "v1/services/aigc/multimodal-generation/generation";
+    }
+    return `v1/services/aigc/text-generation/generation`;
+  },
 };
 
 export const Tencent = {
@@ -570,6 +575,9 @@ const alibabaModes = [
   "qwen-max-0403",
   "qwen-max-0107",
   "qwen-max-longcontext",
+  "qwen-omni-turbo",
+  "qwen-vl-plus",
+  "qwen-vl-max",
 ];
 
 const tencentModels = [

app/utils/chat.ts

@@ -3,7 +3,7 @@ import {
   UPLOAD_URL,
   REQUEST_TIMEOUT_MS,
 } from "@/app/constant";
-import { RequestMessage } from "@/app/client/api";
+import { MultimodalContent, RequestMessage } from "@/app/client/api";
 import Locale from "@/app/locales";
 import {
   EventStreamContentType,
@@ -70,8 +70,9 @@ export function compressImage(file: Blob, maxSize: number): Promise<string> {
   });
 }
 
-export async function preProcessImageContent(
+export async function preProcessImageContentBase(
   content: RequestMessage["content"],
+  transformImageUrl: (url: string) => Promise<{ [key: string]: any }>,
 ) {
   if (typeof content === "string") {
     return content;
@@ -81,7 +82,7 @@ export async function preProcessImageContent(
     if (part?.type == "image_url" && part?.image_url?.url) {
       try {
         const url = await cacheImageToBase64Image(part?.image_url?.url);
-        result.push({ type: part.type, image_url: { url } });
+        result.push(await transformImageUrl(url));
       } catch (error) {
         console.error("Error processing image URL:", error);
       }
@@ -92,6 +93,23 @@ export async function preProcessImageContent(
   return result;
 }
 
+export async function preProcessImageContent(
+  content: RequestMessage["content"],
+) {
+  return preProcessImageContentBase(content, async (url) => ({
+    type: "image_url",
+    image_url: { url },
+  })) as Promise<MultimodalContent[] | string>;
+}
+
+export async function preProcessImageContentForAlibabaDashScope(
+  content: RequestMessage["content"],
+) {
+  return preProcessImageContentBase(content, async (url) => ({
+    image: url,
+  }));
+}
+
 const imageCaches: Record<string, string> = {};
 export function cacheImageToBase64Image(imageUrl: string) {
   if (imageUrl.includes(CACHE_URL_PREFIX)) {