MemoryDrivenChatbot

Web-based AI assistant that chats naturally with users while subtly learning and storing their personal information—specifically, name, age, and gender—over time.

Features

• Conversational AI powered by OpenAI Chat API
• User info (name, age, gender) extraction & storage via Supabase
• Persistent user ID using localStorage
• Typing animation and paginated chat history
• Modern React UI with Tailwind CSS
• Supabase Edge Functions for backend logic
• GPT token/cost saving measures:
  • Limit message count to 10.
  • Skips user extraction prompt (function tool), if all user info is present.
• Additional Debug Tools in UI:
  • Info Button in top-right - Fetches and displays the user info that has been extracted so far.
  • X Button - Resets user id in local-storage - Start new chat session

Screenshot

Setup Instructions

1. Clone repository: https://github.com/chandan-m/MemoryDrivenChatbot.git
2. Update API keys and URLs in the .env files (See next section)
3. Start Backend
   • cd supabase
   • Local:

     supabase start
     supabase migration up
     supabase functions serve chat (Optional: To see logs in terminal)
     

     To seed data:

     supabase db reset
     supabase db seed --file ./supabase/seed.sql
     

   • Deployment:

     supabase secrets set --env-file ./supabase/functions/.env
     supabase functions deploy chat
     

4. Start Frontend (local)

   cd frontend
   npm install
   npm run dev
   

Env Files

Frontend:

frontend/.env

VITE_SUPABASE_URL=http://127.0.0.1:54321
VITE_SUPABASE_ANON_KEY=<anon-key>

frontend/.env.production

VITE_SUPABASE_URL=<supabase-domain>
VITE_SUPABASE_ANON_KEY=<anon-key>

Backend:

supabase/functions/.env

OPENAI_API_KEY=<api-key>

Model Configuration (Optional)

Update the following model parameters (if required) in supabase/functions/chat/constants.ts:
Defaults:

• GPT_MODEL - gpt-4o-mini
• GPT_MODEL_BEHAVIOUR - Prompt to set the behaviour of the AI assistant - You are a helpful assistant that chats naturally with users
• MAX_CHAT_HISTORY_CONTEXT - # of past messages sent as context to the model - ``

Architecture:

Message Flow (High-Level)

[React Frontend]
   ↓ (user message)
[Supabase Edge Function (sendMessage)]
   ├─ If new user, create user record with UUID
   ├─ If `extractUserInfo` flag is true:
   │    └─ Calls OpenAI Function Tool:
   │        "Extract name, age, gender from previous messages"
   │    └─ Updates user with extracted info
   └─ Sends prompt + chat history to OpenAI
      ├─ Store message in DB
      └─ Stores response
   ↓
[Frontend receives reply + updates UI]

Info Extraction Logic

1. Function Tool JSON Schema:

{
  name: "extract_user_info",
  description: "Extract name, age, and gender of the user from the chat history",
  parameters: {
    type: "object",
    properties: {
      name: { type: "string", nullable: true },
      age: { type: "integer", nullable: true },
      gender: { type: "string", enum: ["Male", "Female", "Other"], nullable: true }
    }
  }
}

2. If extractUserInfo flag is passed:
   • The last 5–10 messages are sent to OpenAI with this tool.
   • The returned values are stored via Supabase updateUser() function.
   • Missing fields are set as null.

Detailed Backend Flow

sendMessage(user_id, message)

1. Get or Create User

   • Calls getOrCreateUser(user_id) from dataAccess.
   • If the user ID doesn't exist in DB, creates a new user with a constant UUID (during testing or first use).

2. Fetch Recent Chat History

   • Loads the most recent messages using fetchMessages(user_id, limit) for conversational context.

3. Determine if Info Extraction is Needed

   • Runs hasAllUserInfo(user) to check if name, age, gender are already known.
   • If not, sets a flag: extractUserInfo = true.

4. Call chatCompletionHelper()

   • Constructs a message sequence:

     [
       { role: "system", content: GPT_MODEL_BEHAVIOUR },
       ...chatHistory,
       { role: "user", content: message }
     ]

   • If extractUserInfo is true:
     • Includes tools metadata and tool_choice: "auto" to allow OpenAI to call a function tool.
     • If a function call to extract_user_info is made, it parses arguments to extract:
       • name?: string
       • age?: number
       • gender?: string
     • If no assistant message is returned, a second call is made without tools to generate the reply.

5. Update User Info

   • If new user fields are detected and non-null, updates the DB via updateUser(user_id, userInfo).

6. Save Messages

   • Uses saveMessages() to insert both user and assistant messages into Supabase, with timestamps:

     [
       { user_id, role: "user", content: message },
       { user_id, role: "assistant", content: reply }
     ]

7. Send Response

   • Returns:

     {
       user_id,
       reply: <assistant’s reply>,
       savedMessages: [userMessage, assistantReply]
     }