This repository contains the design documentation and a functional prototype for integrating a Large Language Model (LLM) Copilot into an existing Enterprise Resource Planning (ERP) ecosystem.
🚀 Project Overview The goal of this project is to enhance an existing ERP system with an AI Assistant that serves multiple roles (Sales, Finance, and Admin). The system is designed to scale from 100 to 1,000 users while maintaining high reliability, strict data privacy, and efficient GPU utilization.
Key Features Role-Based Access Control (RBAC): Security is enforced at the backend level, ensuring users can only access data permitted by their organizational role.
Grounded AI (RAG): Uses Retrieval-Augmented Generation to eliminate hallucinations by grounding AI responses in real ERP data.
Scalable Serving: Implementation of 4-bit quantization and request queuing to handle the compute-heavy GPT-OSS 20B model on internal infrastructure.
Streaming UI: Support for real-time text streaming and request cancellation to optimize user experience and compute costs.
🛠️ "Thin Slice" Prototype The included Python implementation demonstrates a core vertical slice of the system:
Secure Data Fetching: A Python-based data layer that filters sensitive columns before they reach the model.
Permission Enforcement: A logic-based check that blocks unauthorized users (e.g., Guest users) from accessing product or sales data.
Context Injection: A simulation of the prompting strategy used to ground the LLM in specific CSV-based "ERP data."
📋 Design Documentation The design_document.md file provides a deep dive into:
Day 1 Discovery Strategy: How to map an undocumented codebase and align with stakeholders.
System Architecture: High-level diagrams of the Frontend, Orchestrator, and Model-Serving layers.
Failure Mode Analysis: Prevention and containment strategies for GPU overload, hallucinations, and prompt injection.
Evaluation Plan: A framework for tracking Faithfulness, Latency, and Refusal Correctness.
⚙️ Setup & Requirements Environment: Python 3.10+
Libraries: pandas, transformers, torch
Hardware: Optimized for Intel Arc GPUs (using IPEX) and standard NVIDIA CUDA environments.
How to use this description: Title: Use "OneICT ERP AI Copilot Integration".
About Section: Use the first paragraph of the "Project Overview".
Topics/Tags: Add LLM, ERP-Integration, RAG, AI-Engineering, Python, System-Design.