Layer 3 — Domain Models · Chapter 2
General Models and Domain Models
General-purpose and domain-specific models are not strict opposites. They form a spectrum. A broad foundation model can be lightly adapted for a field, while a deeply specialized model may still inherit most of its capabilities from a general foundation.
When a general model may be enough
A general model can be effective when the task uses ordinary language, the cost of occasional error is low, current knowledge can be retrieved from trusted sources, and the workflow includes human review. Drafting routine correspondence or summarizing nontechnical documents may not require a new specialized model.
When specialization becomes valuable
- The inputs are not ordinary text: medical images, chemical structures, audio signals, satellite data, or machine telemetry.
- The domain has unusual terminology or reasoning patterns.
- Errors create material financial, legal, medical, or safety consequences.
- The system must perform consistently under a narrow and measurable standard.
- Privacy, latency, or deployment constraints require a smaller local model.
- A company possesses proprietary data that can produce a genuine performance advantage.
| Dimension | General model | Domain model |
|---|---|---|
| Coverage | Broad range of subjects and tasks | Narrower field or data type |
| Vocabulary | Common and cross-domain language | Technical, professional, or local terminology |
| Evaluation | General benchmarks and user preference | Expert-defined domain tests and operational outcomes |
| Cost | Often available through an API | May require data, adaptation, validation, and maintenance |
| Advantage | Flexibility and rapid deployment | Reliability, efficiency, compliance, or unique capability |
Hybrid systems are common
Many strong systems combine both approaches. A general model handles conversation, planning, and explanation; a specialized model analyzes the domain-specific input; retrieval systems supply current facts; and deterministic software enforces business rules.