A new RAG system

GSW - RAG

MY new video explores a novel computational framework, the Generative Semantic Workspace (GSW), designed to overcome the critical limitations of standard Retrieval-Augmented Generation (RAG) in processing long-form, narrative-rich text.

We begin by diagnosing the core problem: RAG’s reliance on retrieving disconnected, semantically similar text chunks leads to context fragmentation, making it nearly impossible for Large Language Models to track the stateful, spatiotemporal evolution of entities over time. This fundamentally hinders their ability to perform true episodic memory tasks.

The GSW ArXiv research preprint (link in my video) addresses this by proposing a neuro-inspired, two-part architecture that constructs a persistent, structured internal model of the narrative world, moving beyond fact retrieval to enable genuine situational understanding.

The first key component of the GSW is the Operator, a semantic parser that processes local text chunks to produce a structured, intermediate representation. This local “semantic graph” is actor-centric, identifying not just entities but also their assigned roles (e.g., ‘suspect’, ‘presenter’) and dynamically evolving states (e.g., ‘captured’, ‘nervous’). The states function as contextual modulators on the action probabilities defined by the roles, formally capturing the nuances of a situation.

The second component is the Reconciler, which acts as a memory integrator. It takes the local semantic graphs from the Operator and merges them into a single, coherent global workspace. The Reconciler performs critical functions like entity co-reference resolution (linking ‘Dr. Thorne’ to ‘he’), state updating across the narrative timeline, and spatiotemporal grounding of events, ensuring a logically consistent and continuously updated world model.

Finally, we show the empirical power and efficiency of this architecture. Instead of overwhelming an LLM with raw, noisy text, the GSW-powered Question-Answering pipeline first identifies relevant entities in a query, then queries its own structured workspace to generate concise, focused narrative summaries.

When evaluated on the Episodic Memory Benchmark (EpBench), this method shows state-of-the-art performance, dramatically improving recall on complex queries that require synthesizing information across up to 17 different document chapters. Crucially, this structured approach reduces the query-time context token count by over 51% compared to the next most efficient baseline, highlighting a tractable path towards building AI systems that are not only more accurate at narrative reasoning but also substantially more efficient and less prone to hallucination.