A decision model is a small classifier that answers yes/no questions. It is faster and costs less than a language model for these questions. When you turn it on, it makes these decisions in place of the language model:
| Decision | When | What it decides |
|---|---|---|
| Page classification | Parsing | Which pages of a source to index, and what each page contains |
| Duplicate checks | Parsing | Which extracted snippets repeat one another |
| Reranking | Each MCP, API, and library search | The order of the results. Applies only when Reranking is on |
The language model still extracts the documentation and writes the snippets. If the decision model fails or times out, the language model makes that decision.
Turn on a decision model#
The decision model is off by default. To turn it on:
Open the AI models settings
In the web UI, go to Settings > AI models.
Turn on the decision model
In the Language model card, turn on Use a decision model.
Choose a provider
Select a provider, enter its API key, and optionally a model and an endpoint. Then click Test decision model.
Save
Click Save and apply. Reranking uses the decision model right away. Parsing uses it from the next parse or refresh.
Settings#
| Setting | Description |
|---|---|
| Provider | The service that runs the decision model. See the table below |
| Model | The decision model name. Leave it blank to use the provider’s default |
| API key | API key for the provider. Stored encrypted in the workspace database |
| Endpoint | Overrides the provider’s URL, for example to go through a proxy. Required for Custom |
Providers#
Decision models are served by several companies. Context7 supports these APIs:
| Provider | API | Default model |
|---|---|---|
| OpenRouter | OpenRouter Decisions API. One key for decision models from many companies | typesafe/jev-1.13-20260917 |
| OpenAI | OpenAI Decisions API | gpt-6-luna |
| Vercel AI Gateway | AI Gateway evaluate API | typesafe-ai/jev-1.13.0 |
| Custom | Any endpoint that implements the System One API, such as TypeSafe or a self-hosted model | jev-1.13 |
Test decision model sends one small duplicate check to the provider. It reports a wrong key, model, or endpoint before you save.
Choosing a model#
Page classification and duplicate checks use thresholds that are tuned on Jev 1.13, the default model for every provider except OpenAI. Another model can index a different set of pages. Test a new model on one library before you refresh all of them.
Data sent to the provider#
When the decision model is on, Context7 sends page text, snippet text, and search queries to the decision model provider. Choose a provider that meets your data policy, or use Custom with a model that runs in your infrastructure.