Completed the Specialty AI Architecture
During repeated runtime verification of the Knowledge Foundation, it became clear that concentrating entity extraction, relation generation, knowledge generation, and update decisions into a single Knowledge Structuring AI caused the prompt to become excessively large. As a result, the LLM increasingly prioritized rules and database schemas over the actual evidence. Various improvements, including schema filtering with SchemaSelector, prompt refinements, and JSON structure redesign, were introduced, but they did not fundamentally solve the problem.
To address this, the architecture was redesigned from the ground up. The monolithic AI was replaced with a specialty-based approach that separates processing by domain such as User, Company, Project, ParentTask, ChildTask, AI, Document, and Code, connected through a shared execution foundation consisting of the GenericCandidateBuilder, CandidateValidator, decision processing, and the Knowledge Repository. Rather than building a separate AI per domain, a shared execution foundation switches behavior based on per-domain definitions and prompts.
The role of the LLM was fundamentally redefined. Instead of generating complete entities, its responsibility is now limited to extracting values from evidence. Those extracted values are assembled into common templates by the application, while validation, decision making, and persistence are handled through a shared, deterministic runtime. This approach minimizes probabilistic output from the LLM while allowing the application to keep entity structures and persistence behavior consistent.
As a result, domain-specific processing now runs on the same shared runtime, reducing prompt size while improving stability, maintainability, and extensibility. When adding a new domain, the goal is to extend the system mainly through its definition and prompt while reusing the shared execution foundation.
The Knowledge Foundation has evolved from a monolithic AI architecture into one with a clear separation of responsibilities between the LLM and the application. The LLM focuses on extracting values from evidence, while the shared runtime assembles those values into common templates and performs validation, decision making, and persistence, aiming for a reproducible runtime that is less dependent on probabilistic LLM output. As a result, the architecture has evolved into a knowledge foundation at the core of the Knowledge OS, better suited to continuous feature expansion and future growth.