LLM providers¶
This project uses OpenRouter as the unified API gateway to access various Large Language Models from different providers. OpenRouter provides a single interface to access models from multiple LLM providers, simplifying integration and offering competitive pricing. The cost is paid by Mira Geoscience up to a given maximum usage per user.
The following table provides a summary of the available model. After extensive testing, the gpt-5.4 model has been selected for use with Petro Lingo. It is the most advanced OpenAI model, best for complex geological interpretation and summarization tasks with zero data retention. Refer to the benchmark, which justifies this selection.
Model |
Provider |
Value Rating |
Data Retention |
Pricing |
|---|---|---|---|---|
gpt-5.4 |
Premium |
Zero |
$1.25 input / $7.50 output |
|
custom |
User-defined |
Variable |
Variable |
Local costs |
Pricing: USD per million tokens
Value Rating: Premium (high cost, high performance), Budget (low cost, good performance)
Data Retention: The duration for which the provider may store portions of your data
About data policy
The application restricts the list of providers from OpenRouter so that your data are never used to train or improve models. The model below is served exclusively through providers with a zero data retention (ZDR) policy: the providers will not store your data for any period of time.
Custom Model Configuration¶
You can configure the application to use your own models from different providers or custom endpoints. To do so, go to Files > Preferences > LLM settings in Geoscience ANALYST.
OpenAI API compatible models¶
URL: The OpenAI-compatible API endpoint from your model provider (e.g.
https://yourmodelprovider.com/api/v1/)Model: The name or identifier of the model you intend to use (e.g.
your-model-name)Enter LLM API key: Click to enter the API key for your model provider
Ollama models (local)¶
Model: The name of the Ollama model you have installed locally.
Important
We cannot guarantee compatibility for custom models, as they might not fully support some API options and their answer format might differ from the expected one.