[AI Tutor] CT-562: Add s3 system prompt

apps/src/aiTutor/chatApi.ts

@@ -40,11 +40,17 @@ const logViolationDetails = (response: OpenaiChatCompletionMessage) => {
 export async function postOpenaiChatCompletion(
   messagesToSend: OpenaiChatCompletionMessage[],
   levelId?: number,
-  tutorType?: AITutorTypesValue
+  tutorType?: AITutorTypesValue,
+  levelInstructions?: string
 ): Promise<OpenaiChatCompletionMessage | null> {
   const payload = levelId
-    ? {levelId: levelId, messages: messagesToSend, type: tutorType}
-    : {messages: messagesToSend, type: tutorType};
+    ? {
+        levelId: levelId,
+        messages: messagesToSend,
+        type: tutorType,
+        levelInstructions,
+      }
+    : {messages: messagesToSend, type: tutorType, levelInstructions};
 
   const response = await HttpClient.post(
     CHAT_COMPLETION_URL,
@@ -57,7 +63,7 @@ export async function postOpenaiChatCompletion(
   if (response.ok) {
     return await response.json();
   } else {
-    return null;
+    throw new Error('Error getting chat completion response');
   }
 }
 
@@ -74,14 +80,13 @@ const formatForChatCompletion = (
  * to `postOpenaiChatCompletion`, then returns the status of the response and assistant message if successful.
  */
 export async function getChatCompletionMessage(
-  systemPrompt: string,
   formattedQuestion: string,
   chatMessages: ChatCompletionMessage[],
   levelId?: number,
-  tutorType?: AITutorTypesValue
+  tutorType?: AITutorTypesValue,
+  levelInstructions?: string
 ): Promise<ChatCompletionResponse> {
   const messagesToSend = [
-    {role: Role.SYSTEM, content: systemPrompt},
     ...formatForChatCompletion(chatMessages),
     {role: Role.USER, content: formattedQuestion},
   ];
@@ -90,7 +95,8 @@ export async function getChatCompletionMessage(
     response = await postOpenaiChatCompletion(
       messagesToSend,
       levelId,
-      tutorType
+      tutorType,
+      levelInstructions
     );
   } catch (error) {
     MetricsReporter.logError({
@@ -100,7 +106,12 @@ export async function getChatCompletionMessage(
     });
   }
 
-  if (!response) return {status: Status.ERROR};
+  if (!response)
+    return {
+      status: Status.ERROR,
+      assistantResponse:
+        'There was an error processing your request. Please try again.',
+    };
 
   switch (response.status) {
     case ShareFilterStatus.Profanity:

apps/src/aiTutor/constants.ts

@@ -3,9 +3,6 @@ import {
   AITutorInteractionStatus as Status,
 } from '@cdo/apps/aiTutor/types';
 
-export const systemPrompt =
-  'You are a tutor in a high school classroom where the students are learning Java using the Code.org curriculum. Answer their questions in plain, easy-to-understand English. Do not write any code. Do not answer the question if it is not about Java or computer programming. Please format all responses in Markdown where appropriate. Use four spaces for indentation instead of tabs.';
-
 // Initial messages we set when the user selects a tutor type.
 // General Chat
 export const generalChatMessage = {

apps/src/aiTutor/redux/aiTutorRedux.ts

@@ -1,7 +1,6 @@
 import _ from 'lodash';
 import {getChatCompletionMessage} from '@cdo/apps/aiTutor/chatApi';
 import {createSlice, PayloadAction, createAsyncThunk} from '@reduxjs/toolkit';
-import {systemPrompt as baseSystemPrompt} from '@cdo/apps/aiTutor/constants';
 import {savePromptAndResponse} from '../interactionsApi';
 import {
   Role,
@@ -20,7 +19,6 @@ export interface AITutorState {
   aiResponse: string | undefined;
   chatMessages: ChatCompletionMessage[];
   isWaitingForChatResponse: boolean;
-  chatMessageError: boolean;
   isChatOpen: boolean;
 }
 
@@ -42,7 +40,6 @@ const initialState: AITutorState = {
   aiResponse: '',
   chatMessages: initialChatMessages,
   isWaitingForChatResponse: false,
-  chatMessageError: false,
   isChatOpen: false,
 };
 
@@ -72,15 +69,8 @@ export const askAITutor = createAsyncThunk(
       scriptId: aiTutorState.aiTutor.scriptId,
     };
 
-    let systemPrompt = baseSystemPrompt;
     const levelInstructions = instructionsState.instructions.longInstructions;
 
-    if (levelInstructions.length > 0) {
-      systemPrompt +=
-        '\n Here are the student instructions for this level: ' +
-        levelInstructions;
-    }
-
     const storedMessages = aiTutorState.aiTutor.chatMessages;
     const newMessage: ChatCompletionMessage = {
       role: Role.USER,
@@ -91,11 +81,11 @@ export const askAITutor = createAsyncThunk(
 
     const formattedQuestion = formatQuestionForAITutor(chatContext);
     const chatApiResponse = await getChatCompletionMessage(
-      systemPrompt,
       formattedQuestion,
       storedMessages,
       levelContext.levelId,
-      chatContext.actionType
+      chatContext.actionType,
+      levelInstructions
     );
     thunkAPI.dispatch(
       updateLastChatMessage({

apps/src/aiTutor/views/AssistantMessage.tsx

@@ -47,6 +47,7 @@ const AssistantMessage: React.FC<AssistantMessageProps> = ({message}) => {
 
   const shouldRenderFeedbackButtons =
     message.id &&
+    message.status !== Status.ERROR &&
     message.status !== Status.PROFANITY_VIOLATION &&
     message.status !== Status.PII_VIOLATION;
 

apps/src/aiTutor/views/UserMessage.tsx

@@ -16,15 +16,13 @@ const PROFANITY_VIOLATION_USER_MESSAGE =
   'This chat has been hidden because it is inappropriate.';
 const PII_VIOLATION_USER_MESSAGE =
   'This chat has been hidden because it contains personal information.';
-const ERROR_USER_MESSAGE =
-  'There was an error getting a response. Please try again.';
 
 const statusToStyleMap = {
   [Status.OK]: style.userMessage,
   [Status.UNKNOWN]: style.userMessage,
   [Status.PROFANITY_VIOLATION]: style.profaneMessage,
   [Status.PII_VIOLATION]: style.piiMessage,
-  [Status.ERROR]: style.errorMessage,
+  [Status.ERROR]: style.userMessage,
 };
 
 const getMessageText = (status: string, chatMessageText: string) => {
@@ -34,7 +32,6 @@ const getMessageText = (status: string, chatMessageText: string) => {
     case Status.PII_VIOLATION:
       return PII_VIOLATION_USER_MESSAGE;
     case Status.ERROR:
-      return ERROR_USER_MESSAGE;
     default:
       return chatMessageText;
   }

dashboard/app/controllers/openai_chat_controller.rb

@@ -1,7 +1,14 @@
 class OpenaiChatController < ApplicationController
+  S3_AI_BUCKET = 'cdo-ai'.freeze
+  S3_TUTOR_SYSTEM_PROMPT_PATH = 'tutor/system_prompt.txt'.freeze
+
   include OpenaiChatHelper
   authorize_resource class: false
 
+  def s3_client
+    @s3_client ||= AWS::S3.create_client
+  end
+
   # POST /openai/chat_completion
   def chat_completion
     unless has_required_messages_param?
@@ -16,16 +23,19 @@ def chat_completion
     # If the content is inappropriate, we skip sending to OpenAI and instead hardcode a warning response on the front-end.
     return render(status: :ok, json: {status: filter_result.type, flagged_content: filter_result.content}) if filter_result
 
+    # The system prompt is stored server-side so we need to prepend it to the student's messages
+    system_prompt = read_file_from_s3(S3_TUTOR_SYSTEM_PROMPT_PATH)
+
+    # Determine if the level is validated and fetch test file contents if it is
+    test_file_contents = ""
     if validated_level?
       level_id = params[:levelId]
       test_file_contents = get_validated_level_test_file_contents(level_id)
-      messages = params[:messages]
-      messages.first["content"] = messages.first["content"] + " The contents of the test file are: #{test_file_contents}"
-      messages.second["content"] = "The student's code is: " + messages.second["content"]
-    else
-      messages = params[:messages]
     end
 
+    updated_system_prompt = add_content_to_system_prompt(system_prompt, params[:levelInstructions], test_file_contents)
+    messages = prepend_system_prompt(updated_system_prompt, params[:messages])
+
     response = OpenaiChatHelper.request_chat_completion(messages)
     chat_completion_return_message = OpenaiChatHelper.get_chat_completion_response_message(response)
     return render(status: chat_completion_return_message[:status], json: chat_completion_return_message[:json])
@@ -39,7 +49,55 @@ def validated_level?
     params[:type].present? && params[:type] == 'validation'
   end
 
-  def get_validated_level_test_file_contents(level_id)
+  def add_content_to_system_prompt(system_prompt, level_instructions, test_file_contents)
+    if level_instructions.present?
+      system_prompt += "\n Here are the student instructions for this level: #{level_instructions}"
+    end
+
+    if test_file_contents.present?
+      system_prompt += "\n The contents of the test file are: #{test_file_contents}"
+    end
+
+    system_prompt
+  end
+
+  private def prepend_system_prompt(system_prompt, messages)
+    system_prompt_message = {
+      content: system_prompt,
+      role: "system"
+    }
+
+    messages.unshift(system_prompt_message)
+    messages
+  end
+
+  private def read_file_from_s3(key_path)
+    full_s3_path = "#{S3_AI_BUCKET}/#{key_path}"
+    cache_key = "s3_file:#{full_s3_path}"
+    unless Rails.env.development?
+      cached_content = CDO.shared_cache.read(cache_key)
+      return cached_content if cached_content.present?
+    end
+
+    if Rails.env.development?
+      local_path = File.join("local-aws", S3_AI_BUCKET, key_path)
+      if File.exist?(local_path)
+        puts "Note: Reading AI prompt from local file: #{key_path}"
+        return File.read(local_path)
+      end
+    end
+
+    # Note: We will hit this codepath in dev if the file is not found locally
+    content = s3_client.get_object(bucket: S3_AI_BUCKET, key: key_path).body.read
+
+    # In production and test, cache the content after fetching it from S3
+    unless Rails.env.development?
+      CDO.shared_cache.write(cache_key, content, expires_in: 1.hour)
+    end
+    return content
+  end
+
+  private def get_validated_level_test_file_contents(level_id)
     level = Level.find(level_id)
 
     unless level