Decision Models
A decision model doesn't generate text: you give it some content and a question, and it answers with a calibrated probability. Questions come in three forms:
- Yes / no — the probability that the answer is "yes" (e.g. "Does this text contain health data?" → 93%).
- Choice — one probability per option in a list (e.g. Invoice 91%, Contract 6%, Other 3%).
- Score — a rating on ordered levels (e.g. from "off-topic" to "answers directly").
These models (Jev, pplx-decider, Microsoft-Decision, D3…) are fast (around 50 to 500 ms per question) and cheap (a few cents per million input tokens, output is free). Since they produce no prose, there is nothing to interpret: the answer is always usable as is, and you can set a confidence threshold.
Where they are used
- Switch — classify the document across the routes, with a minimum score below which the flow takes "Default".
- Interceptors — block processing when the probability passes a threshold (e.g. health data present).
- Annotation filter — drop annotations judged irrelevant.
- RAG profiles — reranking of retrieved passages, instead of an LLM.
- Chat memory — a pre-filter that skips the extraction call when a turn holds nothing worth remembering.
- Benchmark and RAG evaluation — an alternative judge to the default LLM.
Configuration
From Manager → Settings → AI Models (Decision models section):
- Name — label shown in the selectors.
- Provider — chosen among the Decision-type providers configured by the administrator.
- Model name — e.g.
pplx-decider-latest,typesafe/jev-1.13,vllm-d3-mini. The list comes from the provider; manual entry stays possible. - Context window (optional) — bounds the size of the text sent with each call. Empty: about 12,000 characters.
The "Test" button asks the model a real yes/no question and shows the probability obtained and the call duration. The first call to a self-hosted model can take a few seconds while it loads.
Supported providers
They all speak the same exchange format ("System One"). Only the address changes:
| Provider (key) | Models | Endpoint |
|---|---|---|
typesafe | Jev | https://api.typesafe.ai by default |
perplexity | pplx-decider, pplx-decider-flash | https://api.perplexity.ai by default |
openrouter | Jev, Microsoft-Decision, pplx-decider… | OpenRouter's |
ollama (0.35 and later) | local decision models | Ollama's |
openai_compatible | LM Studio, llama.cpp, vLLM (D3, open-weight pplx-decider…) | server address, with /v1 |
Microsoft-Decision is only reachable through OpenRouter for now. With LM Studio, only the models it classifies as decision models are offered: it refuses the others.
Costs
Calls are counted in Organization → Finances, based on input tokens ("per token" pricing) or call duration ("per hour" pricing, suited to a self-hosted model). The rate is set in the Decision model pricing section, and the By decision model breakdown details consumption.
Deletion
A model still used by a pipeline, a RAG profile or an interceptor can only be deleted by picking a replacement model, substituted everywhere. A benchmark or RAG evaluation that used it as a judge simply falls back to the default AI.