Skip to content

Quickstart

pip install "mimir-decisions[local]"

from mimir import Context, Field, Mimir, Passage

model = Mimir.from_pretrained("Mythologic/MIMIR-1")

# The short form: the chosen option id, or None when the model abstains or defers
team = model.pick(
    "Hi, invoice INV-8821 was billed twice. Please refund the duplicate.",
    ["billing", "technical", "sales"],
)  # "billing"

# Context pairs ticket prose with typed structured fields (amounts, IDs, metadata)
context = Context(
    passages=[
        Passage(
            title="Ticket #4091",
            text="Hi, we were billed twice ($2,400 total) on invoice INV-8821. Please refund the $1,200 duplicate today or we will cancel our plan.",
        )
    ],
    fields=Field.from_json({
        "invoice_id": "INV-8821",
        "duplicate_amount": 1200,
        "customer_plan": "enterprise",
    }),
)

# The question is optional: model.choose(context, options) asks "Which option applies to this context?"
result = model.choose(
    context,
    "Which department should handle this request?",
    options={
        "billing": "Billing: invoices, payments, refunds",
        "technical": "Technical: bugs, outages, system errors",
        "sales": "Sales: pricing, new contracts",
        "other": "Other: everything else",
    },
)

The first call downloads the model from the Hugging Face Hub at the revision this package version pins, verifies the manifest's Sigstore signature, checks every file against the SHA-256 in the manifest, and loads it. Nothing is read until every check passes.

On a machine with a CUDA GPU, device="auto" serves the GPU; pass device="cpu" to stay on the CPU. The same weights serve both.

result.status            # Status.DECIDED, Status.ABSTAINED or Status.DEFERRED
result.answer            # "billing", or None when no option applies
result.probabilities     # calibrated probability of each option id
result.relevant_context  # the parts of the context the answer relied on, most relevant first
result.certificate       # the evidence, when the answer is certified

answer is always what the model thinks. status says whether it cleared the operating floor:

  • DECIDED — act on answer.
  • ABSTAINED — no listed option applies (clean OOD rejection).
  • DEFERRED — the answer came in below the floor (threshold or certified mode); have a person review it. The answer is still there.

The Risk Dial: Choosing Your Operating SLA

The release certifies thresholds across four finite-sample risk levels (model.info().risk_levels = (0.005, 0.01, 0.02, 0.05)):

  • risk=0.01 (99.0% SLA) — Mission-critical aerospace / financial SLA: Extremely conservative statistical floor. Only near-certain answers execute automatically; anything uncertain is deferred for human review.
  • risk=0.05 (95.0% SLA, the default) — High-throughput web agent / customer workflow: The optimal operating point for autonomous agents, unlocking high automated throughput while maintaining mathematical error guarantees on held-out data.
  • mode="standard" — Returns the raw calibrated argmax without finite-sample risk deferral; risk then only selects the certificate attached to the result. mode="certified" also defers below the threshold certified at risk.

A ticket's real text decides; the same options over a one-line summary can come back ABSTAINED. Pairing prose with typed fields (Field.from_json) provides the strongest signal for the decision engine. Decisions covers the question and option shapes that decide.

Next

  • Decisions — yes/no questions, claim verification, rankings, ratings, and estimates, over passages, tables, and JSON.
  • The certificate — what the risk level guarantees, and how to certify thresholds on your own data.
  • Python — batching, async, offline use, and the HTTP client.