The Challenge
Integrating existing ERP and Shop Schema
With popular ERP solutions having large amounts of varying data structures, the initial challenge is to identify a Model Schema that can be structured and flexible at the same time. Table descriptions have to be fine tuned and understandable to the AI Agent, while also retaining modification possibilities.
Consistency and Accuracy
LLM hallucinations and inconsistency in SQL generation initially produced inaccurate results, a challenge that had to be overcome fast as accuracy was a main goal of the technical feasibility discovery.
The Solution
In close alignment with procelo tosca, Gradion's approach came into play. It included:
IS-Status Analysis
The company has increased its operating speed fivefold, making everything from managing orders to delivering products much faster and more efficient.
AI Agent
Preparing the AI Agent and Interface for intuitive data interactions.
Migrating Data
Migrating anonymized data to prototype SQL DB, supported by secondary Vector DB.
Fine Tuning
Targeting accuracy by enhancing the agents' step-by-step progress through vector examples, table descriptions, and prompting.
The final prototype is able to follow semantic changes to query specific differences such as orders versus purchases, redact results - and most importantly build the SQL queries with a high level of accuracy to present users with an easy to use analysis tool.
The Impact
The outcome was a portable prototype ready to be showcased to potential real users and move towards product development.
- 8-Week Prototype Delivery
- 80%+ Result accuracy achieved
Product Goals
- Simplify Onboarding
- Lower Cost
- Target SME
