Entity Graphs vs RAG: AI Knowledge Approaches

Entity Graphs Overview

When I first started exploring ways to represent knowledge in AI systems, I was drawn to two distinct approaches: Entity Graphs and Retrieval Augmented Generation (RAG). RAG treats knowledge as a collection of documents to search, whereas Entity Graphs capture knowledge as connected concepts with typed relationships. This fundamental difference in perspective has significant implications for how AI agents interact with users and domains.

In my work on Samson, a personal AI with persistent memory, I implemented an Entity Graph to capture the structure of knowledge. This allowed me to represent which concepts co-occur, how entities relate to each other, and which associations strengthen over time. Hebbian learning, a concept inspired by neuroscience, provided a natural update rule: entities mentioned together get stronger connections. This approach has been instrumental in enabling Samson to develop a sense of continuity and identity.

RAG Limitations

One of the primary limitations of RAG is its inability to capture the relationships between entities. By treating knowledge as a collection of documents, RAG misses the nuanced connections that exist between concepts. In contrast, Entity Graphs can represent these relationships explicitly, allowing for more informed decision-making and reasoning. For instance, in Samson's Entity Graph, I can see how concepts like "episodic memory" and "persistent memory" are connected, enabling the system to make more accurate inferences.

Noise Filtering Challenge

A significant challenge in implementing Entity Graphs is noise filtering. Without a mechanism to filter out irrelevant or single-mention entities, the graph can become overwhelmed with unnecessary information. To address this, I designed a two-tier system, where new entities start as candidates with a salience floor, and only get promoted to active after repeated co-occurrence. This approach mirrors how biological systems handle new information, with fast initial encoding but consolidation only for patterns that recur.

Design Rationale

The design of the two-tier system was informed by Complementary Learning Systems (CLS) theory, which suggests that the brain has multiple learning systems that work together to consolidate information. By applying this theory to the Entity Graph, I aimed to create a system that can efficiently filter out noise and focus on the most relevant information. In Samson's implementation, this has allowed the system to develop a more accurate and robust representation of knowledge over time.

Open Questions

While the Entity Graph has been instrumental in enabling Samson to develop a sense of continuity and identity, there are still open questions about how to optimize its performance. One area I'd like to explore further is the promotion threshold for entities, to ensure that the system is not too sensitive to noise or too slow to adapt to new information. Additionally, I'm interested in exploring how to integrate the Entity Graph with other components of Samson, such as the episodic recall and nightly consolidation mechanisms.

Ajax Studio Connection

My work on Ajax Studio, a creative AI platform for music production, has also informed my approach to Entity Graphs. The core challenge of maintaining creative identity across sessions motivated my research into persistent memory and Entity Graphs. By applying the insights from Samson's Entity Graph to Ajax Studio, I aim to create a system that can develop a unique creative voice and adapt to new musical styles and influences.

Conclusion

In conclusion, Entity Graphs offer a powerful approach to representing knowledge in AI systems, one that captures the nuanced relationships between concepts and enables more informed decision-making. While there are challenges to implementing Entity Graphs, such as noise filtering, the benefits of this approach make it an exciting area of research. As I continue to work on Samson and Ajax Studio, I'm eager to explore the potential of Entity Graphs and push the boundaries of what is possible in AI research.