A developer this week open-sourced a toolset that compiles scattered database documentation (DDL statements, column descriptions, business rules) into structured "knowledge packs," letting AI Agents look up business data like flipping through a dictionary. The core conclusion stings: in a tested enterprise database, 48 of 61 business rules (78.7%) couldn't be automatically linked to specific tables—because these rules never clearly stated which fields they applied to.
What this is
It's called the OKF Compiler, short for Open Knowledge Format. Think of it this way: enterprise knowledge lives in chaos—DDL SQL sits in system A, column descriptions live in Excel, and business rules live in senior employees' heads. The compiler organizes all of this into structured Markdown file sets, complete with table-of-contents indexes and bidirectional links. Once an AI Agent receives these files, it reads the "menu" first, loading details on demand rather than stuffing all documentation into the context window at once. This isn't a large-model upgrade—it's infrastructure work to let AI reliably read your database.
Industry view
What deserves our attention here isn't the technical detail but a judgment: the bottleneck for enterprise AI Agent deployment isn't that models aren't smart enough—it's that knowledge engineering hasn't caught up. The open-source author repeatedly stresses one principle—"the compiler reports gaps, doesn't guess to fill them"—which is essentially an admission of an awkward fact: when AI can't find context, it makes things up.
Counterarguments exist. Some view this as over-engineering: today's mainstream large models have context windows reaching a million tokens—just stuff the database schema in directly. The author's rebuttal: long context doesn't equal accurate retrieval. As token costs explode, the risk of AI confidently hallucinating amplifies. A deeper concern: with 78.7% of rules failing to auto-link, companies still need dedicated staff to maintain this knowledge base in the short term. The "AI replaces data analysts" narrative will have to wait.
Impact on regular people
For enterprise IT: Buying a large model is just the start. Budgeting and headcount for "knowledge organization" will be hard costs over the next two years.
For individual careers: People who understand both business and data are more valuable in the AI era. What's scarce isn't the AI, but the intermediaries who translate business language into structured knowledge.
For consumer markets: In the short term, "AI auto-generating financial reports and answering business questions" remains mostly at the PPT demo stage, far from stable production deployment.