Retail Conversational Information System

Large retail organizations run on internal knowledge that’s often scattered across intranets, handbooks, and institutional memory — hard for employees to search and harder still to keep current. I consulted for a large German retail company to build a conversational in-house information system that let staff simply ask questions and get answers, using the open-source Rasa framework.

The Challenge

Internal information chatbots live or die on the quality of their training data and dialog design — get the intents and entities wrong and the bot either misunderstands routine questions or, worse, answers confidently with the wrong information. The system also had to fit an enterprise environment: reliable, maintainable by the client’s own team going forward, and grounded in the company’s actual internal knowledge rather than generic conversational patterns.

My Approach

I led the implementation using Rasa, at the time one of the leading open-source frameworks for building task-oriented dialog systems, and personally led the data generation effort — designing the intents, entities, and example dialogs the system needed to understand real employee questions rather than idealized ones. Because Rasa is a framework rather than an out-of-the-box product, a large part of the engineering work was in shaping the dialog design and training data specifically around the client’s internal processes and terminology.

Outcome

The result was a working internal assistant that let employees get answers to common questions conversationally rather than digging through documentation. The project is also a good example of a pattern that shows up across my consulting work: the modeling framework is rarely the hard part — the data generation and dialog design around it is where a system succeeds or fails.

Technologies: Rasa, task-oriented dialog design, data generation, intent/entity modeling.