PoC for internal AI chatbot
ewb has an extensive internal knowledge database at its disposal for answering customer inquiries—however, the search for relevant information is still carried out manually. For customer service employees, this is very time-consuming, and the quality of the results also depends heavily on individual experience and knowledge.
How could an internal AI chatbot improve this? Together with ewb, we developed a proof of concept (PoC) for this.
THE CHALLENGES
The challenge was not a lack of information, but rather its efficient use. ewb's knowledge base is extremely extensive, heterogeneous in content, and structured in different ways.
Customer service employees must therefore identify currently relevant content, interpret it correctly, and translate it into understandable, customer-oriented answers. This process takes time and leads to varying results depending on the level of experience.
The planned internal chatbot was designed to understand the content of inquiries, identify relevant knowledge content, and generate high-quality response suggestions. The goal of our project was to realistically evaluate the technical feasibility and concrete added value of such an approach.
THE CHALLENGES
The challenge was not a lack of information, but rather its efficient use. ewb’s knowledge base is extremely extensive, heterogeneous in content, and structured in different ways.
Customer service employees must therefore identify currently relevant content, interpret it correctly, and translate it into understandable, customer-oriented answers. This process takes time and leads to varying results depending on the level of experience.
The planned internal chatbot was designed to understand the content of inquiries, identify relevant knowledge content, and generate high-quality response suggestions. The goal of our project was to realistically evaluate the technical feasibility and concrete added value of such an approach.
OUR APPROACH
We relied on an isolated PoC based on a retrieval-augmented generation architecture.
To do this, we divided the existing knowledge content into text segments, generated semantic embeddings, and stored them in a vector database. At the same time, we also represented user queries semantically in order to identify text passages with matching content in a targeted manner.
We then made the relevant content available to a generative language model for structured and comprehensible response generation.
Our focus was deliberately on the quality of retrieval and the controlled use of existing content. However, we decided not to train our own model. The PoC was implemented locally and without access to productive systems.
OUR APPROACH
We relied on an isolated PoC based on a retrieval-augmented generation architecture.
To do this, we divided the existing knowledge content into text segments, generated semantic embeddings, and stored them in a vector database. At the same time, we also represented user queries semantically in order to identify text passages with matching content in a targeted manner.
We then made the relevant content available to a generative language model for structured and comprehensible response generation.
Our focus was deliberately on the quality of retrieval and the controlled use of existing content. However, we decided not to train our own model. The PoC was implemented locally and without access to productive systems.
THE RESULT
The PoC clearly showed that a chatbot can reliably identify relevant content and compile it in an understandable way. However, the quality of the responses depends heavily on the structure, consistency, and timeliness of the underlying documents. Through iterative test runs and targeted adjustments to text splitting and the pipeline, we achieved measurable improvements.
The chatbot can support customer service, especially with complex or rare questions and when onboarding new employees.
A productive introduction can unlock additional efficiency potential. However, this requires clear governance regulations, a well-thought-out operating concept, and clean data protection integration.
THE RESULT
The PoC clearly showed that a chatbot can reliably identify relevant content and compile it in an understandable way. However, the quality of the responses depends heavily on the structure, consistency, and timeliness of the underlying documents. Through iterative test runs and targeted adjustments to text splitting and the pipeline, we achieved measurable improvements.
The chatbot can support customer service, especially with complex or rare questions and when onboarding new employees.
A productive introduction can unlock additional efficiency potential. However, this requires clear governance regulations, a well-thought-out operating concept, and clean data protection integration.
ABOUT EWB
Energie Wasser Bern (EWB) is an independent public-law company owned by the City of Bern and one of the five largest municipal energy suppliers in Switzerland, with around 680 employees.
Its customers include around 70,000 households, 8,000 small and medium-sized enterprises, and 100 large customers. The company supplies the city of Bern and surrounding communities with electricity, district heating, natural gas, biogas, and water, converts waste into energy, and offers services in the field of electromobility and self-consumption solutions.
