Skip to main content

Choosing and switching models

SimplAI is deliberately model-agnostic: anywhere a model is used — an Agent's reasoning, a Tool's LLM step, a Knowledge Base's embeddings, a Voice Agent's STT/TTS — you pick the provider and model from a dropdown, as long as a key for that provider is connected. This page is practical guidance on which to pick and how to change it later.

Pick by capability tier, not brand​

Rather than defaulting to one vendor, match the model to the job:

  • Frontier / high-reasoning models — complex multi-step reasoning, planning agents, nuanced instructions, hard extraction. Highest quality, highest cost and latency.
  • Balanced mid-tier models — most production agents and tools. Good quality at materially lower cost; usually the right default.
  • Small / fast models — high-volume, latency-sensitive, or simple classification/routing steps where a large model is overkill.

A common pattern is to mix tiers within one application: a small model for routing or guardrail-style checks, a mid-tier model for the main response, and a frontier model only for the hardest step.

Embedding models: keep them consistent​

For Knowledge Bases, the embedding model used at ingestion must match the one used at query time — you can't mix embeddings within a KB. Choose the embedding model when you create the KB and keep it stable; changing it means re-ingesting. See Knowledge & Data.

Re-ranking is optional but high-leverage​

Re-ranking reorders retrieved chunks by relevance after the vector search. On larger or noisier Knowledge Bases it often improves answer quality more than swapping the LLM. Enable it in the KB's retrieval settings once a re-ranking provider key is connected. See Re-ranking.

Switching models later​

Because the model is a configuration choice, you can switch at any time:

  1. Open the Agent, Tool step, or Voice Agent configuration.
  2. Pick a different provider/model in the model dropdown (the provider's key must be connected).
  3. Re-test before publishing — different models follow instructions differently, so re-run your evaluations. See Evaluation.

When switching, watch for prompt portability: a base instruction tuned for one model family may need light adjustment for another.

Bring Your Own Key (BYOK) and fallback​

  • BYOK — connect your own provider keys at workspace or application level so usage bills to your provider account. See the provider pages under LLMs.
  • Fallback — for resilience, configure fallbacks where supported (for example, voice agents support fallback transcribers and fallback voices) so a provider hiccup degrades gracefully instead of failing the run.