Jay Persona Simulator 🤖

This project is a multi-stage AI Discord bot designed to scrape the chat history of a specific target user, chunk it, index it using a custom Vector Search, and simulate that user's persona in live chat using Gemini.

🚀 Features Implemented

1. High-Performance Discord Scraping

Bypassed Discord's standard 10,000-message offset pagination limit by utilizing the messages/search API endpoint with chronological min_id sliding. This allows lightning-fast scraping of all historical messages sent by a target user across an entire server, safely appending to a .jsonl file in real-time.

2. Conversational Chunking Strategies

Developed multiple strategies to group raw messages into contextual "chunks" suitable for AI ingestion:

• 15-Minute Conversational Chunks: Groups messages sent in the same channel within a 15-minute window.
• 2-Hour Windows: Groups messages into strict 2-hour blocks.
• Daily Windows: Groups messages by UTC calendar day.
• 2-Day Windows: Groups messages into 48-hour blocks.

3. Keyword TF-IDF Analysis & Local Vector Database

• Built an internal analytical engine using tiktoken to count tokens and analyze keyword viability.
• Created a Local Sparse Vector Database (retriever.js) built on TF-IDF scoring. It extracts the top 15 most defining keywords for every conversational chunk, stores them as metadata, and uses an Inverted Index for lightning-fast RAG (Retrieval-Augmented Generation) lookups.

4. AI Query Expansion & Context Analysis

Implemented a "two-stage" AI pipeline:

1. Query Expander (Gemini 2.5 Flash): When the bot receives a prompt, it feeds the last 100 messages to Gemini 2.5 Flash via a JSON-structured mini-prompt. The AI extrapolates related keywords, synonyms, and evaluates temporal context (e.g., expanding "this term" into "course, study, term, class").
2. Database Search: These expanded keywords are passed to the Local Vector Database to pull the absolute most relevant historical conversation chunks.

5. Gemini 3.0 Persona Generation

The final RAG context, the temporal analysis, up to 100 random baseline persona quotes, and the full chat history are packaged into a massive prompt and sent to gemini-3.0-flash. The model mimics the target user's syntax, lowercase habits, slang, and specific knowledge to formulate a hyper-realistic response.

6. Local Testing UI Simulator

Built an Express/HTML local web simulator (server.js) that mimics Discord's interface.

• Run the simulator to tweak generation parameters (Temperature, Top P, Top K Context chunks, Persona Examples) in real-time.
• Features a Debug Panel to visualize exactly what keywords the Query Expander extracted and what RAG chunks were retrieved.

🛠️ How to Run

Local Testing Web Simulator

To test the bot, tweak parameters, and view debug context in your browser:

node server.js

Then navigate to http://localhost:3000.

Live Discord Bot

To run the live bot in your Discord server (listening for @mentions):

npm start

(Or run node index.js)

Prerequisites

• Node.js (v18+)
• A Discord Bot Token (set as DISCORD_TOKEN in .env)
• A Google Gemini API Key (set as GEMINI_API_KEY in .env)