Hartmann Group, a pioneer in healthcare and hygiene products since 1818, faced a crucial challenge in refining their GenAI chatbot, Eva. With wound care products in over 100 countries, Hartmann aimed to use GenAI for improved customer support. Eva should answer customer’s questions on how to use Hartmann’s products correctly to treat wounds. Unfortunately, Eva's initial performance was not consistent, nor accurate. The guardrails implemented to ensure only correct answers would be given resulted in 50% of the questions being ignored...  Eva required significant improvements to meet the operational standards and expectations for handling complex healthcare queries in a highly regulated environment.

That is where addData stepped in. We tackled Hartmann's challenges by auditing the existing implementation to enhance Eva's performance.  

Our approach focused on improving data pré-processing and enhancing contextual understanding, allowing the chatbot to provide accurate and reliable responses.  

1. Audit Report

Our first step was to conduct a thorough audit of Eva's performance. We identified six key areas where the current setup was falling short.  

2. Develop an Improvement Plan

Based on the audit findings, we formulated a phased improvement plan with three critical action points aimed at enhancing Eva's framework.

3. Implementation and Continuous Refinement

The improvement plan was executed in several phases, each designed to progressively enhance the chatbot’s hit rate and functionality:

- Phase 1: Redefining Scope and Initial Improvements

We began by refining the scope of the chatbot , focusing on relevant documents and web content. This phase significantly boosted the hit rate to 80%.

- Phase 2: Integrating Document-Level Context

By incorporating document-level context for PDFs and section-level context for web pages, we further refined the chatbot’s accuracy, raising the hit rate to 85%.

- Phase 3: Comprehensive Contextual Enrichment

We enhanced the chatbot’s responses by adding section-level context for both web pages and PDFs. This comprehensive enrichment process increased the hit rate to 88%.

- Phase 4: Reducing Noise and Enhancing Context

We refined the scope and reduced noise in the knowledge database, which led to a significant increase in the hit rate. Additionally, we enriched the context with metadata, overlapping chunks, and structured the data more effectively. These improvements, along with enhanced prompts, elevated the hit rate to an impressive 98%

‍

Our collaboration not only improved customer engagement but also significantly boosted operational efficiency, aligning with Hartmann's rigorous standards for quality and compliance.

Enhanced Accuracy and Reliability

The hit rate increased from an initial 50% to an industry-leading 98%, ensuring Eva provided reliable and accurate responses.

Operational Efficiency

Eva was transformed into a fully operational chatbot capable of delivering precise information in a regulated environment.

Improved Customer Satisfaction

The improvements led to significantly higher customer satisfaction and engagement, reinforcing Hartmann’s reputation for innovative digital solutions.

‍

"With the guidance of addData, we transformed our chatbot into a fully operational agent, seamlessly integrating Hartmann data and LLM comprehension, resulting in a B2C GenAI chatbot that answers absolutely accurate responses in a highly regulated environment.”

Tamas Balog, Global Digital Marketing and Customer Engagement Lead at Hartmann