Combining LLMs with Graph Databases

Lessons from Klarna

One theme of this newsletter is that 1) LLMs aren't good at everything, and 2) most products you use (like ChatGPT) are actually a bundle of LLMs and augments that work together to deliver an effective product.

One unlikely, but potentially fantastic pairing, are LLMs and Knowledge Graphs (KGs). Unlike LLMs, which are associative in nature, KGs explicitly add semantic metadata to the corpus, creating structured data available for computation. KGs have existed for years but are notoriously difficult to implement. Now, LLMs may simplify the creation of ontologies for unstructured data, complementing their associative nature with explicit structure.

Klarna aggressively implemented ChatGPT last year but encountered significant challenges. Their “fragmented corporate data landscape confused the LLM”, prompting them to undertake the massive task of cleaning and structuring their information.

Their exploration revealed crucial insights:

Which data was truly valuable? What information was duplicative, incorrect, or contradictory? The root problem was clear - our knowledge was scattered across various SaaS platforms, each with different concepts and structures, creating an unnavigable web that required extensive Klarna-specific expertise to utilize effectively.

Their solution? Consolidate everything using a Knowledge Graph rather than relying on various SaaS applications for primary data storage with basic features layered on top. Once they rationalized their data, business operations became significantly more efficient.

The results were impressive - serious productivity gains emerged as they allowed their internal AI to leverage this knowledge. With help from Cursor AI, they quickly deployed new interfaces and interactions.

Perhaps the future of AI-powered business will center on maintaining clean data, with a single views across accounts and ground truths maintained by background agents. Users can query as needed, while creating and maintaining dashboards becomes straightforward. This would invert the traditional SaaS model as companies bring their technology stacks back in-house.

Looking at business trends that oscillate between consolidation and fragmentation, SaaS has moved IT out of businesses and into centrally managed clouds, and AI may reverse this trend through the data channel. However, extending this reversal to consumers by providing everyone with a personal server would require something far more radical.