You choose an AI model for two jobs in ExactFind: writing answers and building the index. Each job has its own platform and model menu, so you can use the same one for both or mix them.
Go to ExactFind > Settings > Basic and look under AI provider.
Answers and indexing
| Job | What it does | What matters most |
|---|---|---|
| Answers | Picks articles and writes the reply for every question | Speed and cost, because it runs on each question |
| Indexing | Summarises each item when you build the index | Quality, because it only runs when content changes |
Platforms
You can choose Claude (Anthropic), ChatGPT (OpenAI) or an OpenAI-compatible service. Each platform needs its own API key, and the key field for a platform shows once you've chosen it for answers or indexing.
Using two platforms means two accounts to pay and watch. See Set an AI API Spending Limit With Your Provider.
OpenAI-compatible services
This option connects to services such as Grok, OpenRouter, Mistral and Groq. Enter:
- Service base URL, such as
https://api.mistral.ai/v1. It has to use HTTPS, unless the service is on your local network. - Service name, which visitors see in the disclaimer and the suggested privacy wording.
- Your API key for that service.
- The model ID, typed exactly as the service lists it.
The model must support tool calling. Test connection checks this by asking the model to call a small tool. A model that can't fails with "This model can't call tools, which ExactFind needs. Choose a model that supports tool calling." Choose another model from the same service.
We haven't tested models on these services, so try your own questions before you go live. No cost shows when you build the index, because ExactFind doesn't know their prices.
What "(tested)" means
We run each model through a set of test questions, and "(tested)" marks the ones that passed for that job. The label depends on the menu, so a model can be tested for answers, for indexing, or for both.
| Job | Tested models |
|---|---|
| Answers | Claude Haiku 5.5, Claude Opus 5.5, GPT-6.1 Terra, GPT-6 Luna |
| Indexing | Claude Sonnet 5.5, Claude Opus 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra |
You can pick a model without the label, but we haven't confirmed how it behaves for that job. For Claude and ChatGPT, models come from the menu only.
Prices
For Claude and ChatGPT, the line under each model menu shows its price per million tokens, in US dollars. ExactFind refreshes these prices weekly for its estimates. Your provider's bill is the final figure.
After you change something
- Click Save Changes.
- Click Test connection. It tests your saved settings, for both answers and indexing. If it fails, see API Key Not Working? Fix a Failed Test Connection.
- If you changed a model, a notice shows its prices and the Limits tab opens. Check your token limits still suit them.
Changing the answers model takes effect straight away. In ExactFind Knowledge, open chats start again when you change it.
Changing the indexing model doesn't rewrite your index. Only new and changed items use the new model. To redo an entry, use Regenerate. See Edit an Index Entry, Regenerate It or Remove It.