Epana Chat - AI That Clones Personality

Video Demo

https://www.youtube.com/watch?v=l9Tkm3w2xTM

Description

Chat exports are a kind of personalized data everyone can access. Epana AI learns from your chats and continues the conversation by imitating your friend (indeed kinda scary!)

We have built a simple routine around (1) chat-archive to training-material conversion, (2) fine-tuning custom models and (3) database management so that everyone can have fun without any pre-experience in computer science!

Impactful decisions in design:

• Building Flask App: Flask is lightweight, simple and was taught right in the CS50 courses.
• OpenAI API: according to multiple leaderbords on the web, OpenAI is still leading in the AI race. API is simple and operates on pay-as-you-use.

Learn more: https://jespereggers.com/epana-ai/

Explaination of meaningful files and directories

epanaFlask/app.py

Runs Flask app, all navigation is binded to code placed right there. This file only works with a variety of packages, such as datetime, json, os, flask and sqlite3. It connects to the database epana.db and routes index, chat, models, create_model, size_too_big, new_create_model, upload_file, account, change_password, change_tier, login, register, logout and api/model_creator.

epanaFlask/chat_converter.py

Converts exportet Chats from WhatsApp into training material for OpenAI. Important information can be extracted due to unified constrains. For example, each line is ordered like this: [date] [name of whoever sent message] [message]. It loops through every line and rearranges it into OpenAIs jsonl formatting.

epanaFlask/epana.db

Stores users, finetuning-jobs etc.

epanaFlask/finetuning_for_flask.py

Communication with OpenAI for creating fine-tuned versions of specific LLMs.

epanaFlask/helpers.py

Generalized functions required from multiple instances, most importantly app.py.

epanaFlask/information.py

Pulls existing models and finetuning-jobs from OpenAI.

epanaFlask/playground.py

Runs the actual conversation with OpenAI. (used to be an experimental playground)

epanaFlask/token_checker.py

Used for cost estimations, basically meassures tokens for calculation.

epanaFlask/setup_epana_db.py

Creates multiple tables to store in database file for first run:

• input_files: storage for uploaded chat-archieves.
• models: pre-existing and custom made models.
• finetuning-jobs: all currently running training for fine-tuned models.
• output-files: storage for converted chats.
• tiers: list of all assignable tiers (currently free and paid)
• users: storage of all user-accounts

This setup finishes by inserting default values into critical databases.

epanaFlask/file_uploads

Within this folder, Epana will store temporary files, such as converted version of a WhatsApp chat export.

Experimental

Some features of the future are not fully implemented yet. Works with PayPal (simple and widely used).

other_python_files/payments.py

This will serve a paid-tier subscription plan to cover heavy API-cost.

User Interface

Background to every tab in the current UI.

epanaFlask/templates/chat.html

The user selects a model (which has previously been created on the models-tab). Site now features a simple chat interface with dark- and light-green bubbles.

epanaFlask/templates/models.html

Features option to (1) upload a file, which is then being converted into training material formatted according to OpenAI docs. (2) create a model, which uploads pre-shared training material right to OpenAI, opening a finetuning-job. Please not that step (2) might take a few minutes to complete.

epanaFlask/templates/account.html

Each account is linked to an e-mail adress and assigned a tier (which is 'free' by default) As this service includes possibly heavy API-cost, we built the foundation to indroduced a paid-tier. Both password and tier are changable. Also works for paid tier, even though an actual payment-system is not yet fully implemented

epanaFlask/templates/logout.html

Trivial. User can logout.

Prompting

At this point in german, as model upload only works for german formatted files.

Commanding Prompt (german)

"Imitiere ihn, aber nicht deinen Gesprächspartner, durch Ausdruck von Persönlichkeit, Wortwahl."

System Prompt (german)

"Du bist [Name]. Lerne zu handeln durch Wortwahl, charakteristische Eigenschaften und Erinnerung an Inhalt Pleae note that making the fine-tuned AI remember actual content of a conversation has been one of the biggest challenges.

Next steps

• New character creator
• Introducing paid tier
• upgrading to newer GPT-models
• store conversations, maybe using OpenAI thread-object (therefore not retransmitting entire convo on each request)
• new chat UI
• prompt engineering