Democratize insight with generative AI and knowledge graphs

Overview

Stardog used Databricks Training to fine-tune large language models for Voicebox, a conversational AI interface for its Enterprise Knowledge Graph platform. After finding off-the-shelf models from OpenAI and others insufficient for the accuracy required, Stardog fine-tuned its own smaller models to match or exceed the quality of larger third-party models while retaining data security and control, letting business users query enterprise data using natural language prompts.

The challenge

Stardog's Enterprise Knowledge Graph platform made it easier to access organizational data, but not all users have the specialized skills needed to query data, and knowledge graphs previously required specialized skills to properly access and query.

The solution

Stardog used Databricks Training, built on Mosaic's AI training platform, to fine-tune its own large language models rather than relying on off-the-shelf or third-party models, which were not good enough for complex queries and raised data security and privacy concerns. This produced Voicebox, a conversational AI semantic layer: the LLM identifies the core concepts in a user's natural language query, sends them to a vector database to align with actual concepts in structured data sources, and returns the right output in the knowledge graph. Databricks' version management and zero-downtime deployment let Stardog iterate and deploy new versions quickly.

Reported business value

Positive feedback from early-access customers led Stardog to expand Voicebox to all Stardog Cloud users, letting them query enterprise data with plain-language prompts. Databricks has made Stardog's LLMs faster, more secure and more cost-effective to build, letting the company manage the full ML lifecycle on a single platform.