Challenge
The use of large language models in the context of sensitive practice data carries clear risks. LLMs can make incorrect assumptions, provide inaccurate answers, or generate faulty SQL statements.
This can not only lead to incorrect results but also jeopardize database performance, data protection, and access security. Soxes’ task was therefore to develop a system that makes sensible use of AI while ensuring security, consistency, and traceability at all times.
Approach
We proceeded step by step in collaboration with the client. Close coordination ensured that technical implementation and real-world requirements aligned.
Step 1: Understand the Data Set
In a dataset workshop, soxes analyzed existing data structures, typical query scenarios, and potential optimizations. This clarified which information is relevant for the AI and how it can be used securely.
Step 2: Evaluate AI models
Various models were tested for accuracy, stability, and efficiency. The goal was a setup that reliably understands natural language and generates technically appropriate queries.
Step 3: Build the API framework
soxes developed a modular backend with an API that is secure, scalable, and flexibly expandable. This allows the frontend, chat interface, and database to communicate with each other in a controlled manner.
Step 4: Make answers transparent
Using Explainable AI, the processing steps were made transparent. This ensures that it remains clear how a question is interpreted, translated into a query, and output as a response.
Step 5: Continuously improving quality
Through feedback loops with the client, the solution was validated against real-world usage scenarios and specifically optimized. This allowed the quality of the responses to be improved step by step.
Solution
The result is a backend system with an API that translates natural language into structured and validated database queries. When a user asks a question about practice data, the system analyzes the request, generates an optimized query, and verifies it. The result is then not only output in technical terms but also translated into an understandable answer. The solution can be integrated into various chat interfaces, user interfaces, or existing practice software environments.
Result
The proof of value demonstrated that AI can make complex practice data accessible quickly, securely, and in a user-friendly manner. The solution was subsequently fully implemented. For users, this means:
- Answers in seconds instead of time-consuming database searches
- No unverified or risky SQL queries
- Transparent AI results
- faster access to relevant practice data
- a stable foundation for additional AI features
This provided the ERP provider with a productive foundation for meaningfully integrating AI into its own practice software and making data access significantly easier for users.
Technical aspects
The solution is based on a modern Python stack with Django and PostgreSQL in the backend. Celery, RabbitMQ, and Redis were used to ensure scalable operations.
