FAQ

Is this a generative LLM?

No. The models score typed questions and return probabilities. They do not generate prose, explanations or chat turns, and there is no text parser between the model and the answer.

What does noul mean?

noul is a yes/no question. Its result is the probability of true, plus the normalized confidence and metadata. The name avoids overloading “boolean” for a model score that is not a hard logical value.

What is the difference between choice, noul and score?

• choice: pick one option from a list. The core has a multi-label flag, but no model in the current catalog declares multi-label support.
• noul: estimate the probability of yes versus no.
• score: choose an ordered level such as calm, annoyed, furious; the expected level is available separately.

All three are validated against the selected model's capabilities.

Why probabilities instead of labels?

A threshold can turn a probability into a label, but the probability preserves uncertainty and lets each application choose its own operating point. Use classify(..., threshold=...) when you want a label plus a below_threshold flag.

Are the probabilities calibrated? Calibration evidence is measured per model and variant in Benchmarks; do not assume that a probability from one model is directly comparable with a probability from another.

Does it run offline?

Yes, after the model is in the local cache. Only dtk models download needs the network. The library, CLI and REST server otherwise run locally. See Models.

Does it download models implicitly?

No. If a model or variant is missing, the call fails instead of downloading it. Run dtk models download <id> explicitly. This keeps startup deterministic and makes offline operation possible.

Do I need a GPU?

No. The supported runtime is ONNX Runtime on CPU. GPU acceleration is not part of the documented path. Larger models need more RAM and take longer.

Which model should I start with?

For general choice, yes/no and score decisions, start with gliner2.5-multi and its default variant. If your state is not English, start with laya-multilingual. Read Models and the model cards, then measure on a small labelled sample from your domain.

Which variant should I choose?

Start with the catalog default. Try fp32 when quality matters more than latency, and q4/q4f16/q8 when memory or disk is tight. Verify quality on your own data before switching. The default is a sensible starting point, not a universal recommendation.

Can I download several variants of the same model?

Yes. Variants coexist in the cache. Downloading one does not replace another; dtk models use selects the persistent default, and --variant selects one for a single command.

Can I load two variants at the same time?

Yes, for example over REST with gliner2.5-multi@fp16 and gliner2.5-multi@fp32. They occupy separate registry slots. The library can also hold separate DecisionModel instances. Do not assume their probability outputs are interchangeable without evaluation.

Can I mix models in one application?

You can load different models in separate instances, but a decision is made by one model at a time. Different models have different adapters, calibration and behavior; do not average or compare their probabilities without measuring what that means for your task.

Does REST use the TypeSafe and OpenAI schemas?

Yes. The server exposes System One at /v1/systemone and OpenAI Decisions at /v1/decisions, with the pinned request and response shapes documented in REST. Local extensions are namespaced.

How do I select a quantization in REST?

Use a model alias in the standard model field:

{"model": "gliner2.5-multi@fp32"}

There is no variant field in a System One or OpenAI request body. The base id uses the server's --variant or the persisted default.

Is the REST server authenticated?

It can be. dtk serve --api-key ... requires a bearer token, and DECISION_TOOLKIT_API_KEY is an optional environment fallback. Without either, the server is unauthenticated. It is local-first infrastructure: bind it to localhost or put your own authenticated proxy/TLS in front of it when exposed.

Does the server store my states or decisions?

No, not by default. It evaluates a request and returns a response; it does not create a decision database. State and question text are not used as metric labels. Application logs and your own proxy may still retain request metadata; review Operations.

Can I use a custom ONNX model?

Yes, with an explicit adapter and file map, using DecisionModel.from_onnx(...) or dtk models import. Custom models are not catalogued, do not receive hash verification or support, and do not get a guessed layout. See Core API and CLI.

What licenses do the models have?

The toolkit is MIT licensed. Each model has its own license, recorded in its model card; check the card and the upstream model page before redistribution. See THIRD_PARTY_NOTICES.

What Python versions are supported?

Python 3.12 and 3.13. The package requires Python 3.12 or newer.

Is this ready for production?

The 0.4.x line is usable, but still evolving. Pin the version you deploy, read the changelog, and use the operational limits and runbook in Operations. The REST contracts are pinned; the core API can still change between minor releases.

Does it support images or assets?

The core has typed support for assets and box, but every model currently in the catalog is text-only. The REST server rejects image/asset and box requests for this catalog. See REST.

Is it a vector database or search engine?

No. It returns decisions for typed questions. It does not store vectors, index documents or provide a search UI.

How do I report a bug?

Include the exact command or request, the output of dtk doctor --json, the toolkit version, the model and variant, and any error code. Redact state text that you cannot share. Use the repository issue tracker linked from the home page.