feat: use chat completions API

concord-consortium · Dec 3, 2024 · b16c9cc · b16c9cc
1 parent de304c9
commit b16c9cc
Show file tree

Hide file tree

Showing 3 changed files with 61 additions and 57 deletions.
diff --git a/src/components/App.tsx b/src/components/App.tsx
@@ -60,7 +60,7 @@ export const App = observer(() => {
 
   const handleChatInputSubmit = async (messageText: string) => {
     transcriptStore.addMessage(USER_SPEAKER, messageText);
-    assistantStore.handleMessageSubmit(messageText);
+    await assistantStore.handleMessageSubmit(transcriptStore.messages);
   };
 
   return (

diff --git a/src/models/assistant-model.ts b/src/models/assistant-model.ts
@@ -1,7 +1,6 @@
 import { types, flow } from "mobx-state-tree";
 import { initLlmConnection, getTools } from "../utils/llm-utils";
 import { transcriptStore } from "./chat-transcript-model";
-import { Message } from "openai/resources/beta/threads/messages";
 import { getAttributeList, getDataContext } from "@concord-consortium/codap-plugin-api";
 import { DAVAI_SPEAKER } from "../constants";
 import { createGraph } from "../utils/codap-utils";
@@ -17,70 +16,60 @@ const AssistantModel = types
     const initialize = flow(function* () {
       try {
         const tools = getTools();
-        const assistantInstructions =
-          "You are DAVAI, an Data Analysis through Voice and Artificial Intelligence partner. You are an intermediary for a user who is blind who wants to interact with data tables in a data analysis app named CODAP.";
-        const newAssistant = yield davai.beta.assistants.create({
-          instructions: assistantInstructions,
+        const newAssistant = yield davai.chat.completions.create({
           model: "gpt-4o-mini",
+          messages: [
+            {
+              role: "system",
+              content: [{
+                type: "text",
+                text: "You are DAVAI, an Data Analysis through Voice and Artificial Intelligence partner. You are an intermediary for a user who is blind who wants to interact with data tables in a data analysis app named CODAP."
+              }]
+            }
+          ],
           tools,
         });
 
         self.assistant = newAssistant;
-        self.thread = yield davai.beta.threads.create();
       } catch (err) {
         console.error("Failed to initialize assistant:", err);
       }
     });
 
-    const handleMessageSubmit = flow(function* (messageText) {
+    const handleMessageSubmit = flow(function* (messages) {
       try {
-        yield davai.beta.threads.messages.create(self.thread.id, {
-          role: "user",
-          content: messageText,
-        });
-        yield startRun();
+        const tools = getTools();
 
-      } catch (err) {
-        console.error("Failed to handle message submit:", err);
-      }
-    });
+        // Transform messages for the API request.
+        const transformedMessages = messages.map((message: any) => ({
+          role: message.speaker === "DAVAI" ? "assistant" : "user",
+          content: [{ type: "text", text: message.content }],
+        }));
 
-    const startRun = flow(function* () {
-      try {
-        const run = yield davai.beta.threads.runs.create(self.thread.id, {
-          assistant_id: self.assistant.id,
+        const response = yield davai.chat.completions.create({
+          model: "gpt-4o-mini",
+          tools,
+          messages: transformedMessages,
         });
 
-        // Wait for run completion and handle responses
-        let runState = yield davai.beta.threads.runs.retrieve(self.thread.id, run.id);
-        while (runState.status !== "completed" && runState.status !== "requires_action") {
-          runState = yield davai.beta.threads.runs.retrieve(self.thread.id, run.id);
-        }
-
-        if (runState.status === "requires_action") {
-          yield handleRequiredAction(runState, run.id);
+        if (response?.choices[0]?.finish_reason === "tool_calls") {
+          yield handleRequiredAction(response?.choices[0]?.message.tool_calls, messages);
+        } else {
+          transcriptStore.addMessage(
+            DAVAI_SPEAKER,
+            response?.choices[0]?.message.content || "Error processing request."
+          );
         }
-
-        // Get the last assistant message from the messages array
-        const messages = yield davai.beta.threads.messages.list(self.thread.id);
-        const lastMessageForRun = messages.data.filter(
-          (msg: Message) => msg.run_id === run.id && msg.role === "assistant"
-        ).pop();
-
-        transcriptStore.addMessage(
-          DAVAI_SPEAKER,
-          lastMessageForRun?.content[0]?.text?.value || "Error processing request."
-        );
       } catch (err) {
-        console.error("Failed to complete run:", err);
+        console.error("Failed to handle message submit:", err);
       }
     });
 
-    const handleRequiredAction = flow(function* (runState, runId) {
+    const handleRequiredAction = flow(function* (tool_calls, messages) {
       try {
-        const toolOutputs = runState.required_action?.submit_tool_outputs.tool_calls
+        const toolOutputs = tool_calls
           ? yield Promise.all(
-            runState.required_action.submit_tool_outputs.tool_calls.map(async (toolCall: any) => {
+            tool_calls.map(async (toolCall: any) => {
               if (toolCall.function.name === "get_attributes") {
                 const { dataset } = JSON.parse(toolCall.function.arguments);
                 // getting the root collection won't always work. what if a user wants the attributes
@@ -98,22 +87,37 @@ const AssistantModel = types
           : [];
 
         if (toolOutputs) {
-          davai.beta.threads.runs.submitToolOutputsStream(
-            self.thread.id, runId, { tool_outputs: toolOutputs }
-          );
+          const tools = getTools();
 
-          const threadMessageList = yield davai.beta.threads.messages.list(self.thread.id);
-          const threadMessages = threadMessageList.data.map((msg: any) => ({
-            role: msg.role,
-            content: msg.content[0].text.value,
+          const newMessage = toolOutputs.map((toolOutput: any) => ({
+            role: "system",
+            content: [
+              { type: "text", text: toolOutput.output },
+            ],
           }));
 
-          yield davai.chat.completions.create({
-            model: "gpt-4o-mini",
-            messages: [
-              ...threadMessages
+          // Transform existing messages for the API request. They need to be included in the request to maintain context.
+          const transformedMessages = messages.map((message: any) => ({
+            role: message.speaker === "DAVAI" ? "assistant" : "user",
+            content: [
+              {
+                type: "text",
+                text: message.content,
+              },
             ],
+          }));
+          transformedMessages.push(...newMessage);
+
+          const response = yield davai.chat.completions.create({
+            model: "gpt-4o-mini",
+            messages: transformedMessages,
+            tools,
           });
+
+          transcriptStore.addMessage(
+            DAVAI_SPEAKER,
+            response?.choices[0]?.message.content || "Error processing request."
+          );
         }
       } catch (err) {
         console.error(err);

diff --git a/src/utils/openai-utils.ts b/src/utils/openai-utils.ts
@@ -1,5 +1,5 @@
 import { OpenAI } from "openai";
-import { AssistantTool } from "openai/resources/beta/assistants";
+import { ChatCompletionTool } from "openai/resources";
 
 export const newOpenAI = () => {
   return new OpenAI({
@@ -10,7 +10,7 @@ export const newOpenAI = () => {
   });
 };
 
-export const openAiTools: AssistantTool[] = [
+export const openAiTools: ChatCompletionTool[] = [
   {
     type: "function",
     function: {