Deep Dive: Building Reliable Enterprise Agent Knowledge Bases with LLM Wiki & Google OKF

From RAG to LLM Wiki, the ways agents organize and consume knowledge are diversifying. The author first describes a typical software‑development scenario where hundreds of design documents, API specs, and meeting minutes are chunked, embedded, and placed in a vector store, forming a classic RAG pipeline.

For straightforward factual questions (e.g., “What modules does the system have?”) the pipeline works because semantic search can quickly retrieve relevant fragments. However, for complex exploratory queries such as “How to accurately calculate last quarter’s customer churn rate?” the AI must first understand the definition, calculation scope, data sources, and deprecated fields, turning the task into a knowledge‑exploration problem that classic RAG cannot reliably solve.

The core issue is that RAG merely pulls a batch of possibly relevant raw snippets without providing an understandable, navigable, and maintainable knowledge system. In contrast, a Wiki‑style knowledge base organizes useful knowledge into deterministic assets—markdown files with metadata, links, and Git‑managed versioning.

LLM Wiki (proposed by Andrej Karpathy) adds a continuous knowledge layer between raw materials and consumers. New information is read, organized, and merged into the existing Wiki rather than simply indexed. The process includes updating knowledge units, adding cross‑references, resolving conflicts, and generating navigation indexes—all performed by LLMs.

The author illustrates the workflow with a churn‑rate example: instead of a single snippet, a dedicated Wiki page records the definition, calculation period, required tables and fields, and version history, linking to related customer‑info and data‑table pages. The agent navigates the Wiki hierarchy (index.md → crm/metrics → churn‑rate.md) to assemble an accurate query.

Google’s Open Knowledge Format (OKF) 0.2 defines a lightweight, vendor‑neutral specification for such knowledge bundles. An OKF package consists of a directory of UTF‑8 markdown files, each representing a concept with a YAML front‑matter containing at least a type field. Optional fields include title, description, resource, tags, and timestamp. Links between concepts are expressed with standard markdown links, and an index.md serves as the entry point for agents.

OKF does not replace domain‑specific protocols (e.g., OpenAPI) nor prescribe storage or consumption protocols; it merely ensures interoperability of knowledge bundles across tools and agents.

The article then presents a six‑step agent workflow (source manifest, concept discovery, knowledge writing, relationship creation, navigation generation, audit & release) with a table summarizing each stage, core actions, and artifacts.

Finally, a concrete CRM‑support Wiki use case is described. Knowledge sources include data models, complaint‑handling processes, SLAs, OpenAPI specs, notifications, and historical tickets. Stable concepts are identified, markdown pages are drafted, and a pseudocode pipeline shows how an agent lists concepts, enriches raw data, validates drafts, rebuilds indexes, and publishes after owner review.

Throughout, the author emphasizes that LLM Wiki provides the method for turning raw material into a structured knowledge map, while OKF supplies the format that makes the map portable and interoperable.