cactus-react-native 0.2.3 → 0.2.4

package/src/tools.ts

@@ -56,22 +56,18 @@ export class Tools {
 
 export async function parseAndExecuteTool(result: NativeCompletionResult, tools: Tools): Promise<{toolCalled: boolean, toolName?: string, toolInput?: any, toolOutput?: any}> {
   if (!result.tool_calls || result.tool_calls.length === 0) {
-      // console.log('No tool calls found');
       return {toolCalled: false};
   }
   
   try {
       const toolCall = result.tool_calls[0];
       if (!toolCall) {
-        // console.log('No tool call found');
         return {toolCalled: false};
       }
       const toolName = toolCall.function.name;
       const toolInput = JSON.parse(toolCall.function.arguments);
       
-      // console.log('Calling tool:', toolName, toolInput);
       const toolOutput = await tools.execute(toolName, toolInput);
-      // console.log('Tool called result:', toolOutput);
       
       return {
           toolCalled: true,
@@ -80,7 +76,6 @@ export async function parseAndExecuteTool(result: NativeCompletionResult, tools:
           toolOutput
       };
   } catch (error) {
-      // console.error('Error parsing tool call:', error);
       return {toolCalled: false};
   }
 }

package/src/tts.ts

@@ -31,10 +31,7 @@ export class CactusTTS {
       speakerJsonStr,
       textToSpeak,
     )
-    // This part is simplified. In a real scenario, the tokens from 
-    // the main model would be generated and passed to decodeAudioTokens.
-    // For now, we are assuming a direct path which may not be fully functional
-    // without the main model's token output for TTS.
+    // To-DO: Fix
     const tokens = (await this.context.tokenize(formatted_prompt)).tokens
     return decodeAudioTokens(this.context.id, tokens)
   }

package/src/vlm.ts

@@ -3,15 +3,15 @@ import {
   initMultimodal,
   multimodalCompletion,
   LlamaContext,
+  type ContextParams,
+  type CompletionParams,
+  type CactusOAICompatibleMessage,
+  type NativeCompletionResult,
 } from './index'
-import type {
-  ContextParams,
-  CompletionParams,
-  CactusOAICompatibleMessage,
-  NativeCompletionResult,
-} from './index'
+
 import { Telemetry } from './telemetry'
 import { setCactusToken, getTextCompletion, getVisionCompletion } from './remote'
+import { ConversationHistoryManager } from './chat'
 
 interface CactusVLMReturn {
   vlm: CactusVLM | null
@@ -29,9 +29,11 @@ export type VLMCompletionParams = Omit<CompletionParams, 'prompt'> & {
 
 export class CactusVLM {
   private context: LlamaContext
-  
+  protected conversationHistoryManager: ConversationHistoryManager
+
   private constructor(context: LlamaContext) {
     this.context = context
+    this.conversationHistoryManager = new ConversationHistoryManager()
   }
 
   static async init(
@@ -111,26 +113,47 @@ export class CactusVLM {
     return result;
   }
 
-  private async _handleLocalCompletion(
+  private _handleLocalCompletion = async(
     messages: CactusOAICompatibleMessage[],
     params: VLMCompletionParams,
     callback?: (data: any) => void,
-  ): Promise<NativeCompletionResult> {
+  ): Promise<NativeCompletionResult> => {
+    const { newMessages, requiresReset } =
+      this.conversationHistoryManager.processNewMessages(messages);
+
+    if (requiresReset) {
+      this.context?.rewind();
+      this.conversationHistoryManager.reset();
+    }
+
+    if (newMessages.length === 0) {
+      console.warn('No messages to complete!');
+    }
+
+    let result: NativeCompletionResult;
+
     if (params.images && params.images.length > 0) {
-      const formattedPrompt = await this.context.getFormattedChat(messages)
+      const formattedPrompt = await this.context.getFormattedChat(newMessages)
       const prompt =
         typeof formattedPrompt === 'string'
           ? formattedPrompt
           : formattedPrompt.prompt
-      return await multimodalCompletion(
+      result = await multimodalCompletion(
         this.context.id,
         prompt,
         params.images,
         { ...params, prompt, emit_partial_completion: !!callback },
       )
     } else {
-      return await this.context.completion({ messages, ...params }, callback)
+      result = await this.context.completion({ messages: newMessages, ...params }, callback)
     }
+
+    this.conversationHistoryManager.update(newMessages, {
+      role: 'assistant',
+      content: result.content || result.text,
+    });
+
+    return result;
   }
 
   private async _handleRemoteCompletion(
@@ -181,7 +204,6 @@ export class CactusVLM {
   }
 
   async rewind(): Promise<void> {
-    // @ts-ignore
     return this.context?.rewind()
   }