More than a Search Engine: Using Tools and Structured Outputs to Get the Most Out of Your Chatbot

Many people use LLMs for answering questions and summarising information, but few know that they can do more. This talk looks at how tool use and structured outputs enable a chatbot to not just retrieve but also change and even submit user data. This allows a chatbot to go beyond simple augmented retrieval, becoming more than a type of search engine.

As a practical example, this talk looks at how a chatbot has been used to speed up workflow graph configurations that define business rules for a data capture pipeline. Configuring new graphs manually can be time-consuming, since business rules are often complex, with documents passing through many states and rules causing branching behavior.

The solution uses the OpenAI Agents SDK as well as Pydantic models for simple and reliable configuration using a chatbot interface. The chatbot allows users to modify and create workflows using natural language instead of adding states or nodes individually. This talk is therefore aimed at developers who have basic familiarity with chatbots but want to take them further.