ADR-0001 — Keep the core domain-agnostic

Context

The two ideas behind groundedrag — a hard documentary gate before calling the LLM, and citation verification after generation — are easy to build in a way that quietly couples them to one vertical: the coverage map keyed by a domain-specific taxonomy, chunk metadata shaped around one kind of document, prompt templates that assume a particular question format. That coupling is tempting because it is faster to ship a first working version, but it turns the project into a template for one use case rather than a reusable library.

Decision

gate.py and guardrails.py operate only on three generic types — Document, Chunk, and Answer — and the coverage map is a plain topic → [documents] mapping with no assumptions about what a "topic" is. Everything domain-specific (what counts as a topic, how documents are chunked, what the LLM prompt looks like) lives outside the core, in provider/store implementations or in the caller's own configuration.

Alternatives considered

Ship a vertical-specific starter template (e.g., a pre-built internal IT-helpdesk bot) instead of a general library. Rejected: a template gets copied and forked per use case, which means every bug fix and every improvement to the gate or the citation verifier has to be re-applied by hand across every fork. A library with a stable core interface fixes bugs once for everyone who depends on it.

Bake a default taxonomy into the coverage map (e.g., ship with a predefined set of topic categories) to make onboarding faster. Rejected: any default taxonomy is wrong for most domains, and "wrong but built-in" is worse than "absent and obviously the caller's job," because a wrong default fails silently — a topic that doesn't fit the built-in categories would either be forced into the wrong bucket or silently fall through the gate.

Consequences

• Onboarding a new domain means writing a coverage map and, if needed, a loader for that domain's document format — there is no default to fall back on. This is a deliberate cost: it keeps the gate's behavior predictable rather than "mostly right."
• The reference implementations (Chroma + sentence-transformers + Ollama/ Anthropic) exist to prove the interfaces work end-to-end, not to be the only supported path — any VectorStore, Embedder, or LLMProvider that satisfies the Protocol works.

What would invalidate this

If, after a few real integrations, every single caller ends up writing near-identical coverage maps or loaders, that repetition is a signal to extract a second, optional layer of domain presets on top of the generic core — without removing the domain-agnostic core itself.