Architecture of an Auditable Al Chatbot: Multi-Agent Routing with Open Knowledge Format

When I learned about Open Knowledge Format (OKF), a portable knowledge format built with Markdown and YAML, I knew it could be used for more than just building a chatbot. OKF is designed to be vendor-neutral and independent of any particular agent framework or model provider. It demonstrates datasets, schemas, concepts, references, and connected knowledge through examples like GA4 e-commerce data, Stack Overflow data, Bitcoin data, and retail data.

OKF combines two layers: YAML frontmatter stores queryable fields like type, resource, tags, generation details, verification state, and lifecycle status, while the Markdown body contains the prose, schemas, examples, explanations, and other context that people and language models need to read. This structure allows for a more organized and standardized representation of knowledge compared to treating it as an undifferentiated collection of text.

The chatbot I built with OKF is not just an interface connected to a language model; it's a knowledge-publication and retrieval system that decides which information is allowed to be used in a visitor-facing answer. The chatbot interprets a visitor's question, routes it through an approved catalogue, and generates an answer only from records that satisfy publication rules.

These records undergo strict gatekeeping based on factors such as visibility, stability, and human verification before being used in the answer generation process.

While OKF and Retrieval-augmented Generation (RAG) operate at different layers, they both solve distinct problems in the AI space. OKF provides a structured knowledge representation, while RAG retrieves relevant information from a large dataset to generate responses. OKF adds a layer of control and transparency to the AI's decision-making process, ensuring that only relevant, current, and approved information is used in AI-generated responses.