BaseAiCore

Zero-dependency domain foundation: canonical model identity, machine profile, the Unsupported measurement sentinel, capability IDs, IDs, timestamps and the base error hierarchy.

Status: Phases 1–5 shipped (0.4.1) — the full development plan: measurement, identity, IDs, time, hashing, errors, money, cost, model descriptor, runtime profile, measurement subject, the machine profile with its fingerprint, capability identifiers, an ordered DataClassification, and the optional adapter axis on an execution subject. Every public type is exported from baseaicore.__init__ — see the API reference.

Part of the Local AI Suite.

Install

pip install baseaicore

Quickstart

from baseaicore import ModelIdentity, ProviderKind, normalize_digest

identity = ModelIdentity(
    ProviderKind.OLLAMA,
    "qwen3.5:9b-q8_0",
    normalize_digest("1f3a9c4e2b70a1b2c3d4e5f60718293a4b5c6d7e8f90a1b2c3d4e5f607182930"),
)
print(identity.canonical_id)  # ollama/qwen3.5:9b-q8_0@sha256:1f3a9c4e2b70
print(identity.identity_confidence)  # digest (IdentityConfidence.DIGEST)

A capability ID validates its own syntax and knows its place in the vocabulary's specialization tree — the vocabulary's contents belong to SetSpec, not here:

from baseaicore import CapabilityId

specific = CapabilityId("coding.python")
general = CapabilityId("coding")
print(specific.inherits_from(general))  # True — Python-coding evidence also satisfies "coding"
print(general.inherits_from(specific))  # False — the reverse does not hold

A measurement this machine cannot provide refuses to behave like a number, so the idiom that would fabricate one fails loudly instead of producing a plausible chart:

from baseaicore import UNSUPPORTED, is_supported, supported_values

try:
    UNSUPPORTED or 0  # TypeError — an absent measurement is not 0, and refuses to become one
except TypeError as refusal:
    print(refusal)

is_supported(UNSUPPORTED)  # False
supported_values([12.5, UNSUPPORTED, 9.0])  # [12.5, 9.0], and you know two of three were real

The same rule governs money. A price is a dated, sourced observation with a validity window — not a property of the model — and a model with no price costs unknown, never nothing:

from baseaicore import (
    ModelPricing,
    Money,
    PricingSource,
    TokenRates,
    TokenUsage,
    estimate_cost,
    utc_now,
)

pricing = ModelPricing(
    identity=identity,
    rates=TokenRates(
        currency="USD",
        input_per_million_tokens=Money.from_decimal("USD", "3.00"),
        output_per_million_tokens=Money.from_decimal("USD", "15.00"),
    ),
    source=PricingSource.PROVIDER_PUBLISHED,
    observed_at=utc_now(),
)
usage = TokenUsage(
    input_tokens=1_500_000, output_tokens=200_000, cache_write_tokens=0, cache_read_tokens=0
)

estimate = estimate_cost(usage, pricing, at=utc_now())
print(estimate.total)  # 7.5 USD
print(estimate.pricing_hash)  # which price produced that figure

Store the TokenUsage and the pricing_hash, not the money: when a provider changes its prices — which it does on its own schedule, not yours — history re-costs from the counts instead of being quietly wrong (ADR-0030).

See docs/packages/baseaicore/spec.md §7 for the full API and §20 for the acceptance criteria these examples come from.

Documentation

Project documentation lives under docs/. Start with docs/README.md.

Development

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pre-commit install
pytest -m "not live and not performance"

See CONTRIBUTING.md for the full workflow and SECURITY.md for how to report a vulnerability.

License

Apache-2.0 — see LICENSE.