Blame for code/schemas/inference.py

"""Inference response schema — the contract between model adapters and metrics.

Models return a list of DetectionResult dicts per image (COCO-style).
An empty list means no violations were detected.

An empty list is ambiguous on its own, though: it looks identical whether the
model said "no violations" or produced something we failed to parse. Every
response therefore carries a ``parse_status`` recording which of those happened,
so a silent parse failure cannot be mistaken for a clean negative.
"""

from typing import Literal

from pydantic import BaseModel, Field, field_validator

ParseStatus = Literal["ok", "empty", "parse_error", "format_mismatch", "unrecorded"]
"""How the adapter's raw output was interpreted.

* ``ok`` — parsed cleanly, at least one detection
* ``empty`` — parsed cleanly, the model reported no violations (a real answer)
* ``parse_error`` — the output could not be parsed at all
* ``format_mismatch`` — parsed as structured data, but not in the expected shape
* ``unrecorded`` — the adapter did not say. The default, so that results written
  before parse tracking existed are not silently upgraded to ``ok``; every
  current adapter sets a real status.
"""

FormatSeen = Literal["json", "json_repaired", "yaml", "plain_text", "empty", "other"]
"""The output format actually observed, regardless of what was asked for.

``json_repaired`` means the response was JSON in every respect except that it
would not parse as written — in practice, zero-padded integers like
``"category_id": 03``. Kept distinct from ``json`` because it measures a real
compliance failure, and distinct from ``yaml`` because YAML happens to accept
those literals, which previously made well-formed JSON look like a model
answering in the wrong format entirely.
"""


class DetectionResult(BaseModel):
    """A single detection predicted by a model for one image.

    Follows format:
        {category_id, bbox, score}

    An image with no detections simply has zero ``DetectionResult`` entries.
    """

    category_id: int
    bbox: list[float] = Field(..., min_length=4, max_length=4)
    score: float = Field(ge=0.0, le=1.0)

    @field_validator("bbox")
    @classmethod
    def validate_bbox(cls, v: list[float]) -> list[float]:
        """Ensure bbox has exactly 4 elements [x, y, width, height]."""
        if len(v) != 4:
            raise ValueError("bbox must have exactly 4 elements [x, y, width, height]")
        return v


class InferenceResponse(BaseModel):
    """Wrapper for the full model response for a single image.

    This is the single source of truth for the inference→metrics boundary.
    The ``detections`` list may be empty if no violations were detected.
    """

    image_id: str
    detections: list[DetectionResult] = []
    raw_text: str | None = None
    model_name: str
    latency_ms: float

    # Parse tracking. The defaults describe an adapter that said nothing, so
    # older results stay readable without being credited with a clean parse.
    parse_status: ParseStatus = "unrecorded"
    format_seen: FormatSeen | None = None
    parse_error: str | None = None

    # Geometry tracking. ``bbox`` above is always canonical unit [x, y, w, h];
    # this records the convention it was converted *from*, and what the
    # conversion had to repair. None means the file predates geometry tracking —
    # its boxes were stored exactly as the model emitted them and may not be
    # canonical at all.
    box_convention: str | None = None
    box_repairs: dict[str, int] = {}

    # What this one image cost, as reported by the provider. ``None`` means no
    # cost was reported — a local model, or a provider that does not say — and
    # is deliberately distinct from ``0.0``, which claims the call was free.
    # Never estimated from token counts: a benchmark that publishes cost per
    # image must publish what was actually billed.
    cost_usd: float | None = None

    # Why generation stopped, as the provider reported it. ``"length"`` means
    # the answer was cut off at ``max_tokens`` — a fact about our request, not
    # about the model, and the difference between "this model cannot do the
    # task" and "we did not let it finish". Reasoning models make that
    # distinction load-bearing: they spend the token budget thinking before
    # they emit a single character of answer, so a budget that is merely
    # generous for a normal model leaves them with an empty ``content`` and a
    # results file that reads as total failure. ``None`` means the provider did
    # not say, or the adapter is local.
    finish_reason: str | None = None

    # Optional accounting used by local and hosted multi-pass runs. Older
    # result files remain valid because every field is additive.
    input_tokens: int | None = None
    output_tokens: int | None = None
    context_limit: int | None = None
    context_headroom: int | None = None
    provider_diagnostics: dict = {}
    pass_summaries: list[dict] = []

    @property
    def parsed_cleanly(self) -> bool:
        """Whether the model's output was understood, detections or not."""
        return self.parse_status in ("ok", "empty")


class InferenceAttempt(BaseModel):
    """Auditable record of every model call made for one image."""

    image_id: str
    attempt: int = Field(ge=1)
    status: Literal["ok", "failed"]
    response: InferenceResponse | None = None
    error_type: str | None = None
    error: str | None = None