Open Grid Europe Using chatbots to access company data

Chatbots in enterprise use

Corporate data, such as sales or production data, is often stored in different systems. Finding the right information requires specialised knowledge of the various systems, a great deal of time, and often expertise in order to request the information and make it available. Chatbots can, however, make this structured information easily accessible. Current chatbots such as ChatGPT can interact in natural language and perform complex tasks – including generating content, summarising texts, or even answering complex questions. The underlying Large Language Models (LLMs) can also be quickly and easily integrated into existing cloud infrastructures and applications, using Azure OpenAI, for example. These chatbots cannot, however, answer questions such as “What were November 2023’s sales figures?” or “Which compressor system is currently running in Krummhörn [a municipality on Germany’s north-western coast]?” because they do not have access to structured – and often changeable – company and process data.

Integrating up-to-date data into chatbots

inovex worked with Open Grid Europe GmbH to develop a concept and prototype for a chatbot which has precisely this type of access to company data. As a basis for the chatbot, inovex used the concept of an AI agent. These agents have access to the outside world and can interact with external systems, as well as retrieve and provide information. The behaviour of such agents is specified in the language model, allowing the chatbot to use the AI agents to answer user queries.

As a first step, Open Grid Europe and inovex tested the AI agent concept with an interface for compressor station sensor data in a two-week proof of concept.

From vision to solution

Implementing a complex chatbot which integrates structured business assets and interacts with enterprise systems requires accurate planning and scoping to reduce its complexity. inovex’s proof-of-concept approach therefore consists of the following four phases, each of which was carried out in close cooperation with Open Grid Europe:

Workshop: Evaluation and selection of a promising use case in terms of its complexity and potential implementation risks, followed by the formulation of a solution idea.

Preparatory phase: Technical setup and clarification of details, e.g. setting up an Azure subscription and access to OpenAI.

Implementation phase: Implementation of the solution idea in an iterative, agile process with continuous feedback and adjustments.

Final presentation: Presentation of the solution developed, its benefits, and the lessons learned.

Technological background

The solution implemented consists of three components:

1. An API as an interface to the sensor data of the compressor stations,
2. A chatbot application consisting of the AI agent as well as a frontend and
3. The language models delivered through Azure OpenAI.

The interface to the information from the compressor stations was implemented as a lightweight FastAPI interface. It provides information about the sensor data from the compressor stations. Given the short amount of time available, a system export containing real data was used, rather than the production system itself.

The chatbot application itself was implemented using chainlit, a Python/React library. It provides the basic functions for a state-of-the-art chat interface using LLMs. Building on this, Langchain was used to implement the AI agent. The agent consists of two components:

• The planner, which knows the API endpoints and their specification and develops a plan to respond to the user’s request.
• The controller, which executes each step of the plan, calls the endpoints, and interprets the results.

This system uses an iterative process (React pattern) to answer even complex queries.

Conclusion and future plans

Within a very short time, inovex and Open Grid Europe have succeeded in developing an AI agent which uses a chatbot to make Open Grid Europe company data accessible. Authorised and authenticated Open Grid Europe employees can use the Chainlit web interface to interact with the chatbot and ask the compressor stations complex questions. The AI agent is designed to enable additional data and systems to be quickly integrated and further systems to be connected.