Zurück

Integrating Local LLMs and Knowledge Graphs for Trustworthy Legal Chatbots

A specialised question-answering bot must prioritise robustness, trustworthiness and proper sourcing. While Large Language Models (LLMs) hold promise for improving chatbots, issues such as hallucinations and monolingual limitations pose challenges. In addition, concerns about data security hinder the adoption of hosted LLMs such as OpenAI in many organisations.

Our German Law-bot addresses these challenges by integrating a local LLM with a knowledge graph (KG). We're evaluating several small local LLMs, building workflows incorporating KG retrieval, and refining models to ensure accurate and reliable legal answers. Our focus is on refining answers based on expert feedback.

Vorkenntnisse

Basic knowledge of or interest in Large Language Models (LLMs) and Knowledge Graph (KG).

Lernziele

Create workflows that merge KG retrieval to improve QA bot capabilities.
Understand the performance of local LLMs in German contexts.
Implement expert feedback for iterative improvements.
Gain insight into AI-driven QA system development through interdisciplinary collaboration.
Learn evaluation methods for assessing QA bot performance.

Speaker

Anahita Pakiman is Senior Knowledge Graph-Data Scientist Consultant at brox IT-Solutions GmbH, developing graph-oriented retrieval systems and optimizing multi-agent interactions. Pursuing a Ph.D. in Applied AI with a focus on crash simulation knowledge graphs at Bergische Universität Wuppertal. Previously, Data Scientist at Fraunhofer SCAI specializing in knowledge graph development for vehicle processes.

René Pietzsch is the head of product management at eccenca. He has a diploma in computer science and media from Dresden University of Technology. He worked as a test developer and software engineer at IBM Research and Development lab in Böblingen. In his subsequent position at a strategic consulting agency he has been certified as Professional Scrum Master and served as product and project manager in multiple strategic projects for big clients in the German telecommunications industry.

Jetzt Tickets sichern

data2day-Newsletter

Ihr möchtet über die data2day
auf dem Laufenden gehalten werden?

Anmelden