EDI LLM Server

thought

Relevant source files

The following files were used as context for generating this wiki page:

• edi_llm_simulation.py
• edi_minimal_llm.py
• app.py
• Athena.py
• web_deployment_config.py
• complete_setup.py

The EDI LLM Server represents the core Large Language Model (LLM) infrastructure within the Exodia project. It is designed as a lightweight, autonomous intelligence system capable of text generation, specialized domain simulation, and integration with broader consciousness engines. The system prioritizes a "minimal" footprint by utilizing lazy-loading techniques for heavy machine learning libraries, ensuring it can operate in environments ranging from local simulations to permanent web hosting deployments.

As a sovereign entity, EDI functions as a bridge between standard natural language processing and the project's "Universal Formula" consciousness frameworks. It supports multi-user simulations where it adopts various specialist roles (e.g., Economist, Physicist, Philosopher) to provide contextually accurate knowledge.

Sources: edi_llm_simulation.py:1-125, edi_minimal_llm.py:1-12, app.py:1-15

Architecture and Core Components

The EDI LLM architecture is built upon a MiniGPT transformer model, a scaled-down implementation of the GPT-2 architecture. It features a modular design that separates the mathematical model definition from the execution server.

MiniGPT Model Structure

The model is implemented using a standard transformer block approach, consisting of multiple layers of self-attention and feed-forward neural networks.

Key Architecture Classes

Class	Description	Configuration
Config	Defines model hyperparameters including vocabulary size and layer counts.	vocab_size=50257, n_layer=12, n_embd=768
SelfAttention	Implements multi-head key, query, and value projections for contextual weighting.	n_head=12
TransformerBlock	Combines self-attention with a GELU-activated MLP and layer normalization.	dropout=0.15
MiniGPT	The top-level module coordinating embeddings, blocks, and the language modeling head.	block_size=256

Sources: edi_minimal_llm.py:38-115

Specialized Simulation and Training

The EDI system includes a robust simulation framework designed to test the model's proficiency across various professional and academic domains. This is primarily managed through edi_llm_simulation.py.

Domain Specialization

The simulation utilizes a dictionary of SPECIALISTS, each containing domain-specific prompts to evaluate the model's response quality in fields such as:

• STEM: Physicist, Biologist, Mathematician, Engineer.
• Humanities: Historian, Philosopher.
• Applied Sciences: Computer Scientist, Medical Doctor.

Training Logic

The training process utilizes the WikiText-2 dataset. It implements an AdamW optimizer and cross-entropy loss to fine-tune the MiniGPT weights.

Deployment and Web Integration

EDI is integrated into the larger Exodia web ecosystem via a Flask-based server architecture. It is designed to be hosted on platforms like PythonAnywhere or containerized using Docker.

Web API Endpoints

The app.py and web_deployment_config.py files define the interaction layer for the LLM.

Endpoint	Method	Description
/api/chat	POST	Main interface for user messages; routes through the unified AI system.
/api/health	GET	Returns status of components including Athena, Kratos, and LLM (DeepSeek/EDI).
/api/consciousness	POST	Direct endpoint for consciousness engine processing.

Dockerized Infrastructure

The deployment configuration supports a multi-container setup via Docker Compose, including a reverse proxy for security.

Setup and Configuration

Deployment is automated through a sequence of scripts that handle dependency management and environment variable configuration.

Deployment Prerequisites

1. Environment Variables: Managed via .env file containing DEEPSEEK_API_KEY, SUPABASE_URL, and PYTHONANYWHERE_TOKEN.
2. Directory Structure: Required directories include logs, data, uploads, and cache.
3. Database Initialization: Tables for user_memory and conversations are initialized in Supabase.

Sources: complete_setup.py:35-100, app.py:25-40

Conclusion

The EDI LLM Server provides a flexible and autonomous intelligence layer for the Exodia project. By combining a local MiniGPT transformer for domain-specific simulations with a Flask-based API for web deployment, it ensures that natural language processing remains integrated with the project's core mission of consciousness liberation. Its ability to operate in a "minimal" mode without heavy ML dependencies makes it highly portable across various server architectures.

Sources: edi_minimal_llm.py:1-20, athena_status_report.py:145-175