Back to research

Reliable AI Responses: Retrieval, Tools and Clear Boundaries

Engineering

Retrieval-augmented generation helps models use verified knowledge, but retrieval alone does not guarantee reliable answers. Dependable systems combine source quality, explicit uncertainty and tools for exact facts.

Measured systems for practical intelligence - geisten

Language models can produce plausible language even when evidence is missing. Retrieval gives them relevant documents, but a weak retrieval result, stale source or ambiguous prompt can still lead to an unsupported answer.

Three foundations for reliability

  1. Curated source material. Documents need owners, versioning and access rules. Poor source data remains poor when indexed.
  2. An explicit abstention path. The system should state that evidence is insufficient when retrieved material does not support an answer.
  3. Tools for exact operations. Current prices, stock, appointments and calculations should come from an authorised system or deterministic tool, not from generated text.

Small models in narrow roles

Compact models can be excellent coordinators for retrieval and tools. They do not need to answer every question from their parameters. They need to select the right evidence, invoke the appropriate operation and present a traceable result.

Conclusion

Reliability is an architectural property. Good data, clear boundaries and the right tools make a model useful long before a larger parameter count does.