Ground Your AI

LLMs are fluent — but fluency without constraint is noise. Feed an invoice, a contract, or a form to a model and it will extract the key fields regardless of layout or language. The problem is what comes back.

Define a DSL for the expected result — line items, totals, parties, dates — and instruct the model to respond in that form. A Raku Grammar validates the output, rejects malformed responses, and hands clean structured data to the rest of your pipeline. An inversion-of-control harness can tighten the prompt and retry on failure. The LLM provides the reading; the grammar provides the contract.

Close the Gap

Martin Fowler called it the sweet spot — DSLs that are business-readable, if not business-writeable. The goal is not to eliminate programmers; it is to shrink what he called the Yawning Crevasse of Doom between the people who understand the domain and the people who build the system.

A pricing analyst expressing discount rules. A compliance officer defining validation logic. A content editor writing workflow conditions. Each working in a language shaped to their problem, not the machine's. The grammar enforces correctness; the domain expert provides intent. Raku grammars are classes — they compose, inherit, and evolve alongside the domain they describe.

Focus the Code

Consider a pricing engine. Without a DSL, discount rules live buried in nested conditionals — hard to audit, harder to change when the business shifts. With a DSL the rules become readable declarations:

10% off orders above £500
free shipping above £100
loyalty discount 5% after 12 months

The Grammar parses these rules. The Actions apply them. When the rules change — and they always change — you edit the DSL, not the engine. The implementation is hidden; the intent is visible; the tests stay green.

This is what Fowler means by a semantic layer: a thin, expressive surface that maps cleanly onto domain concepts. Raku's Grammar and Actions classes make that layer a first-class citizen of the language — no framework required, no strings evaluated at runtime, no magic.

Concepts and terminology drawn from Martin Fowler, Domain-Specific Languages (Addison-Wesley, 2010) and martinfowler.com/bliki/DomainSpecificLanguage.