Skip to content

decision-toolkit

Local-first decision models on CPU: typed choice, yes/no (noul) and score questions with a calibrated probability for every answer. One ONNX engine, one adapter per model contract, and three surfaces over the same core: library, CLI and REST.

Co-developed by GIA-UH and Syalia S.R.L..

Quickstart Installation

PyPI Python License: MIT

What it is

  • A library that turns a state and typed questions into probabilities, without generating free text.
  • A CLI (dtk) for decisions, classification, batches and model-cache management.
  • An optional REST server implementing the TypeSafe System One schema and the OpenAI Decisions beta schema.
  • A model catalog of open-weights ONNX builds, pinned by revision and SHA256, with selectable quantizations (variants) on CPU.
  • A local-first toolkit: models are downloaded explicitly, decisions run in process, and there is no hosted service in the middle.

What it is not

  • It is not a generative LLM: it does not produce prose, explanations or chat turns.
  • It is not a rules engine: the questions are typed, but the probabilities come from a model.
  • It is not a vector database or a search service; it returns decisions.
  • It does not download models implicitly. dtk models download is the only command that uses the network.
  • It does not persist your states or decisions by default.

Highlights

  • One ONNX engine (onnxruntime), in process, on CPU.
  • Three surfaces, one implementation: library, CLI and REST.
  • Pinned catalog: revision, per-file SHA256 and ONNX-session verification.
  • Variants coexist: download and compare quantizations without replacing the default.
  • Standard REST selection: use the id@variant model alias; there is no non-standard variant field in request bodies.
  • Offline after download: copy the cache or import a local model.

Install

uv pip install "decision-toolkit[onnx]"          # library, CLI and runtime
uv pip install "decision-toolkit[onnx,server]"   # + REST server

See Installation for source checkouts, offline machines, extras, requirements and where the cache lives.

30-second example

dtk models download gliner2.5-multi
dtk models use gliner2.5-multi

from decision_toolkit import DecisionModel, choice, noul

with DecisionModel("gliner2.5-multi") as model:
    decision = model.decide(
        "Meeting with Acme: 10% discount, sign next week.",
        {
            "cat": choice("What kind of record is this?", {
                "meeting": "a meeting or call",
                "tasks": "things to do",
                "other": "anything else",
            }),
            "urgent": noul("Does it need a reply today?"),
        },
    )

print(decision["cat"].choice, decision["cat"].confidence)
print(decision["urgent"].noul)

The same decision from the CLI:

dtk decide --model gliner2.5-multi \
  --state "Meeting with Acme: 10% discount, sign next week." \
  --q 'cat=choice:meeting|tasks|other' \
  --q 'urgent=noul:Does it need a reply today?' --json

Documentation map

If you want to… Read
Go from install to a first decision Quickstart
Install, verify or work offline Installation
Download, inspect or choose a model Models and the model cache
Compare every catalogued variant Variants and quantization
Use the Python API in depth Library guide
Read the full core contract Core API
Run the examples Examples
Copy a workflow Recipes
Use the command line CLI
Call the REST service REST API
Operate a server Operations
Reproduce latency and parity Benchmarks
Fix a problem Troubleshooting
Check a concept or term FAQ · Glossary
Build or publish a release Release

Status

The current release line is 0.4.x. The core API and the REST contracts are usable, but the project is still evolving; pin the version you deploy and read the changelog before upgrading.

REST compatibility is documented in REST. The TypeSafe System One and OpenAI Decisions shapes are treated as pinned contracts; local extensions are namespaced and documented separately.

License

The toolkit is MIT licensed. Each model keeps its own license; see the model cards and THIRD_PARTY_NOTICES. For bugs and feature requests, use the issue tracker.