clairic-flash / clairic_runtime /clair_schema.py
areneau's picture
Add complete Clairic Transformers package
dc7d8aa verified
Raw History Blame Contribute Delete
22.4 kB
"""Field encoding, request validation, and estimate grids for Clairic.
Scalar types include text, numbers, dates, times, datetimes, and durations.
Lists preserve source order. Estimates use declared numeric or date ranges;
forecasts expand into bounded estimate fields for each requested step.
"""
from __future__ import annotations
import datetime
import difflib
import itertools
import math
import re
from typing import Any
READ_TYPES = ("text", "number", "date", "time", "datetime", "duration")
FIELD_TYPES = READ_TYPES + ("list", "estimate")
LIST_ITEM_TYPES = ("text", "number", "date")
DURATION_UNITS = ("seconds", "minutes", "hours", "days", "weeks", "months", "years")
MAX_LIST_ITEMS = 32
MAX_DECIMALS = 6
MAX_FORECAST_HORIZON = 64
FORECAST_INTERVALS = (
"step",
"minute",
"hour",
"day",
"week",
"month",
"quarter",
"year",
)
DEFAULT_NULL_THRESHOLD = 0.5
MAX_GRID_STEPS = 1_000_000
COMMON_KEYS = ("type", "instructions", "unit")
NUMBER_KEYS = ("integer", "decimals")
ESTIMATE_KEYS = ("min", "max", "scale", "step")
FIELD_KEYS = {
"text": (*COMMON_KEYS, "null_threshold"),
"number": (*COMMON_KEYS, "null_threshold", *NUMBER_KEYS),
"date": (*COMMON_KEYS, "null_threshold"),
"time": (*COMMON_KEYS, "null_threshold"),
"datetime": (*COMMON_KEYS, "null_threshold"),
"duration": (*COMMON_KEYS, "null_threshold", *NUMBER_KEYS),
"list": (*COMMON_KEYS, "items", *NUMBER_KEYS),
"estimate": (*COMMON_KEYS, *ESTIMATE_KEYS),
"forecast": (*COMMON_KEYS, *ESTIMATE_KEYS, "horizon", "interval"),
}
ALL_FIELD_KEYS = tuple(
dict.fromkeys(key for keys in FIELD_KEYS.values() for key in keys)
)
SYSTEM_PREAMBLE = (
"Fill each field from the input with a value of its type. Text copies a span of the input. Numbers, "
"dates, times, datetimes, and durations are read from the input or worked out from it. A list holds "
"every matching value in input order. A field is not stated when the input does not determine it. "
"Estimates predict a value inside the declared range."
)
FIELD_NAME = re.compile(r"^[A-Za-z_][A-Za-z0-9_.\-]{0,63}$")
def format_bound(value: float) -> str:
if float(value).is_integer() and abs(value) < 1e15:
return str(int(value))
return f"{value:.6g}"
def field_line(name: str, spec: dict[str, Any]) -> str:
kind = spec["type"]
unit = str(spec.get("unit") or "").strip()
if kind == "estimate" and is_date_estimate(spec):
head = f"estimate date, {spec['min']} to {spec['max']}"
granularity = date_granularity(spec)
if granularity != "day":
head += f", by {granularity}"
elif spec.get("step") and float(spec["step"]) != 1:
head += f", step {format_bound(float(spec['step']))} days"
elif kind == "estimate":
head = (
f"estimate{(' ' + unit) if unit else ''}, {format_bound(spec['min'])} to "
f"{format_bound(spec['max'])}, {spec.get('scale', 'linear')} scale"
)
if spec.get("step"):
head += f", step {format_bound(float(spec['step']))}"
elif kind == "list":
head = f"list of {spec['items']}{number_options(spec)}{(' (' + unit + ')') if unit else ''}"
elif kind == "duration":
head = f"duration in {duration_unit(spec)}{number_options(spec)}"
else:
head = f"{kind}{number_options(spec) if kind == 'number' else ''}{(' (' + unit + ')') if unit else ''}"
instructions = " ".join(str(spec.get("instructions") or "").split())
return f"{name}: {head}." + (f" {instructions}" if instructions else "")
def number_options(spec: dict[str, Any]) -> str:
if spec.get("integer"):
return ", integer"
if spec.get("decimals") is not None:
return f", {int(spec['decimals'])} decimals"
return ""
def duration_unit(spec: dict[str, Any]) -> str:
return str(spec.get("unit") or "seconds")
def allowed_decimals(spec: dict[str, Any]) -> int | None:
"""Decimal places a number, duration, or number-list field allows (None: any)."""
if spec.get("integer"):
return 0
return None if spec.get("decimals") is None else int(spec["decimals"])
def parse_time(value: Any) -> tuple[int, int, int]:
"""("HH:MM" or "HH:MM:SS") as (hour, minute, second); ValueError otherwise."""
if not isinstance(value, str):
raise ValueError(f"time must be a string, got {value!r}") # noqa: TRY004 - callers catch ValueError
parsed = datetime.time.fromisoformat(value)
if parsed.microsecond or parsed.tzinfo is not None:
raise ValueError(f"time {value!r} must be HH:MM or HH:MM:SS")
return parsed.hour, parsed.minute, parsed.second
def format_time(hour: int, minute: int, second: int) -> str:
return f"{hour:02d}:{minute:02d}:{second:02d}"
def parse_datetime(value: Any) -> tuple[str, tuple[int, int, int]]:
"""("YYYY-MM-DDTHH:MM[:SS]") as (ISO date, (hour, minute, second)); ValueError otherwise."""
if not isinstance(value, str) or "T" not in value:
raise ValueError(f"datetime must be YYYY-MM-DDTHH:MM[:SS], got {value!r}")
day, clock = value.split("T", 1)
return datetime.date.fromisoformat(day).isoformat(), parse_time(clock)
def forecast_instructions(instructions: str, step: int, interval: str) -> str:
"""The instructions for one step of a forecast path."""
what = " ".join(str(instructions or "").split()).rstrip(".") or "the series"
unit = "step" if interval == "step" else interval
return f"Forecast {what} {step} {unit}{'s' if step > 1 else ''} after the last observation."
def check_field_keys(fields: dict[str, Any]) -> None:
"""Every field spec uses only the keys its type takes, so a misspelled option ("decimal",
"interger") is an error instead of a constraint that is silently dropped."""
for name, spec in fields.items():
if not isinstance(spec, dict) or spec.get("type") not in FIELD_KEYS:
continue
kind = spec["type"]
for key in spec:
if key in FIELD_KEYS[kind]:
continue
if key in ALL_FIELD_KEYS:
raise ValueError(f"field {name}: {kind} fields don't take {key!r}")
close = difflib.get_close_matches(
str(key), FIELD_KEYS[kind], n=1, cutoff=0.6
)
hint = (
f"; did you mean {close[0]!r}?"
if close
else f"; {kind} fields take {', '.join(FIELD_KEYS[kind])}"
)
raise ValueError(f"field {name} has unknown key {key!r}{hint}")
def expand_forecasts(
fields: dict[str, Any],
) -> tuple[dict[str, Any], dict[str, list[str]]]:
"""Request fields with every ``forecast`` replaced by ``horizon`` estimate fields named
``name.1`` ... ``name.H`` (same range, scale, step, and unit), plus each forecast's step names."""
expanded: dict[str, Any] = {}
groups: dict[str, list[str]] = {}
for name, spec in fields.items():
if not isinstance(spec, dict) or spec.get("type") != "forecast":
expanded[name] = spec
continue
horizon = spec.get("horizon")
if (
isinstance(horizon, bool)
or not isinstance(horizon, int)
or not 1 <= horizon <= MAX_FORECAST_HORIZON
):
raise ValueError(
f"forecast {name} needs an integer horizon from 1 to {MAX_FORECAST_HORIZON}"
)
interval = spec.get("interval", "step")
if interval not in FORECAST_INTERVALS:
raise ValueError(f"forecast {name} has an unknown interval {interval!r}")
base = {
key: value
for key, value in spec.items()
if key not in ("type", "horizon", "interval", "instructions")
}
if is_date_estimate({"type": "estimate", **base}):
validate_date_range(name, {"type": "estimate", **base}, kind="forecast")
else:
validate_number_range(name, {"type": "estimate", **base}, kind="forecast")
groups[name] = []
for step in range(1, horizon + 1):
step_name = f"{name}.{step}"
if step_name in fields or step_name in expanded:
raise ValueError(
f"forecast {name} step {step_name} collides with another field"
)
expanded[step_name] = {
"type": "estimate",
**base,
"instructions": forecast_instructions(
spec.get("instructions", ""), step, interval
),
}
groups[name].append(step_name)
return expanded, groups
# ---------------------------------------------------------------- estimate transforms
def to_unit(value: float, low: float, high: float, scale: str) -> float:
return min(1.0, max(0.0, raw_unit(value, low, high, scale)))
def raw_unit(value: float, low: float, high: float, scale: str) -> float:
"""Unit position without clamping: below 0 or above 1 when the value is outside the range (a
non-positive value on a log scale is -1)."""
if scale == "log":
if value <= 0:
return -1.0
return (math.log(value) - math.log(low)) / (math.log(high) - math.log(low))
return (value - low) / (high - low)
def from_unit(position: float, low: float, high: float, scale: str) -> float:
position = min(1.0, max(0.0, position))
if scale == "log":
value = math.exp(math.log(low) + position * (math.log(high) - math.log(low)))
else:
value = low + position * (high - low)
return min(high, max(low, value))
def resolution_in_unit(spec: dict[str, Any], resolution: float | None) -> float:
if not resolution:
return 0.0
low, high, scale = estimate_range(spec)
if scale == "log":
return math.log1p(resolution) / (math.log(high) - math.log(low))
return resolution / (high - low)
# ---------------------------------------------------------------- date estimates
def is_date_estimate(spec: dict[str, Any]) -> bool:
return spec.get("type") == "estimate" and isinstance(spec.get("min"), str)
def date_granularity(spec: dict[str, Any]) -> str:
step = spec.get("step")
return step if step in ("month", "year") else "day"
def date_index(value: str, granularity: str = "day") -> float:
"""A date as a number on its grid: day ordinal, months since year 0, or the year."""
date = datetime.date.fromisoformat(value)
if granularity == "month":
return float(date.year * 12 + date.month - 1)
if granularity == "year":
return float(date.year)
return float(date.toordinal())
def index_date(index: float, granularity: str = "day") -> str:
whole = round(index)
if granularity == "month":
year, month = divmod(whole, 12)
return datetime.date(year, month + 1, 1).isoformat()
if granularity == "year":
return datetime.date(whole, 1, 1).isoformat()
return datetime.date.fromordinal(whole).isoformat()
def estimate_range(spec: dict[str, Any]) -> tuple[float, float, str]:
"""Numeric bounds and scale of an estimate; date bounds become grid indexes (linear scale)."""
if is_date_estimate(spec):
granularity = date_granularity(spec)
return (
date_index(spec["min"], granularity),
date_index(spec["max"], granularity),
"linear",
)
return float(spec["min"]), float(spec["max"]), spec.get("scale", "linear")
def estimate_step(spec: dict[str, Any]) -> float:
"""The grid step in the numeric space of ``estimate_range`` (0: continuous). Date estimates always
have a grid: ``step`` days (default 1), or one month or year."""
if is_date_estimate(spec):
return (
1.0 if date_granularity(spec) != "day" else float(spec.get("step") or 1.0)
)
return float(spec.get("step") or 0.0)
def estimate_number(spec: dict[str, Any], value: Any) -> float:
if is_date_estimate(spec):
return date_index(str(value), date_granularity(spec))
return float(value)
def estimate_value(spec: dict[str, Any], number: float) -> float | str:
"""An estimate output in the field's own terms: an ISO date for date estimates."""
if is_date_estimate(spec):
low, high, _ = estimate_range(spec)
return index_date(min(max(number, low), high), date_granularity(spec))
return number
def grid_bounds(step: float, low: float, high: float) -> tuple[float, float]:
"""The smallest and largest multiples of ``step`` inside [low, high]."""
first = math.ceil(low / step - 1e-9) * step
last = math.floor(high / step + 1e-9) * step
return float(f"{first:.12g}"), float(f"{last:.12g}")
def snap(value: float, step: float, low: float, high: float) -> float:
"""Nearest grid value inside the range."""
first, last = grid_bounds(step, low, high)
return float(f"{min(last, max(first, round(value / step) * step)):.12g}")
def snap_quantile(value: float, step: float, low: float, high: float) -> float:
"""Quantile of the discrete distribution a continuous one implies when every grid value v owns
[v - step/2, v + step/2): the smallest v with v + step/2 >= the continuous quantile."""
first, last = grid_bounds(step, low, high)
return float(
f"{min(last, max(first, math.ceil((value - step / 2) / step - 1e-9) * step)):.12g}"
)
def step_interval(spec: dict[str, Any], value: float) -> tuple[float, float]:
"""The value-space interval a grid value owns, clipped to the declared range (``value`` in the
numeric space of ``estimate_range``)."""
step = estimate_step(spec)
low, high, _ = estimate_range(spec)
if step <= 0:
return value, value
return max(low, value - step / 2), min(high, value + step / 2)
# ---------------------------------------------------------------- text spans
def _cuts_word(text: str, start: int, end: int) -> bool:
"""Whether text[start:end] splits an ASCII word or number at either edge ("dock" inside "docks",
"12" inside "312"). Other scripts are never treated as cut, so CJK matches are unaffected."""
def joined(left: str, right: str) -> bool:
return left.isascii() and left.isalnum() and right.isascii() and right.isalnum()
return (start > 0 and joined(text[start - 1], text[start])) or (
end < len(text) and joined(text[end - 1], text[end])
)
def _prefer_whole(
text: str, spans: list[tuple[int, int]], limit: int
) -> list[tuple[int, int]]:
whole = [span for span in spans if not _cuts_word(text, *span)]
return (whole or spans)[:limit]
def find_spans(text: str, value: str, limit: int = 16) -> list[tuple[int, int]]:
"""Character spans of ``value`` in ``text``: exact matches, else case-insensitive, else with any
whitespace between words. Within each stage, matches on word boundaries are preferred; substring
matches inside a longer word are kept only when there is no other match."""
value = value.strip()
if not value:
return []
scan = max(limit * 8, 64)
spans: list[tuple[int, int]] = []
start = text.find(value)
while start >= 0 and len(spans) < scan:
spans.append((start, start + len(value)))
start = text.find(value, start + 1)
if spans:
return _prefer_whole(text, spans, limit)
lowered, target = text.lower(), value.lower()
if len(lowered) == len(text):
start = lowered.find(target)
while start >= 0 and len(spans) < scan:
spans.append((start, start + len(value)))
start = lowered.find(target, start + 1)
if spans:
return _prefer_whole(text, spans, limit)
pieces = value.split()
if not pieces:
return []
pattern = re.compile(
r"\s+".join(re.escape(piece) for piece in pieces), re.IGNORECASE
)
spans = [
(match.start(), match.end())
for match in itertools.islice(pattern.finditer(text), scan)
]
return _prefer_whole(text, spans, limit)
# ---------------------------------------------------------------- records
def validate_fields(fields: Any) -> None:
"""A schema with any number of fields (at least one): valid names, known types, and for estimates
a valid range, scale, and step. Requests are checked with this before prediction, so every output
can be held to its field's type."""
if not isinstance(fields, dict) or not fields:
raise ValueError("fields must declare at least one field")
for name, spec in fields.items():
if not isinstance(name, str) or not FIELD_NAME.match(name):
raise ValueError(f"invalid field name {name!r}")
kind = spec.get("type") if isinstance(spec, dict) else None
if kind not in FIELD_TYPES:
raise ValueError(f"field {name} has unknown type {kind!r}")
if "null_threshold" in spec:
threshold = spec["null_threshold"]
if kind in ("estimate", "list"):
raise ValueError(f"{kind} {name} does not take null_threshold")
if (
isinstance(threshold, bool)
or not isinstance(threshold, (int, float))
or not 0 < threshold <= 1
):
raise ValueError(f"field {name} needs null_threshold in (0, 1]")
if kind == "list" and spec.get("items") not in LIST_ITEM_TYPES:
raise ValueError(f"list {name} needs items in {LIST_ITEM_TYPES}")
if kind == "duration" and duration_unit(spec) not in DURATION_UNITS:
raise ValueError(f"duration {name} needs a unit in {DURATION_UNITS}")
numeric = kind in ("number", "duration") or (
kind == "list" and spec.get("items") == "number"
)
if not numeric and ("integer" in spec or "decimals" in spec):
raise ValueError(
f"field {name}: integer and decimals apply to numbers and durations only"
)
if "integer" in spec and not isinstance(spec["integer"], bool):
raise ValueError(f"field {name}: integer must be true or false")
if "decimals" in spec:
decimals = spec["decimals"]
if (
isinstance(decimals, bool)
or not isinstance(decimals, int)
or not 0 <= decimals <= MAX_DECIMALS
):
raise ValueError(
f"field {name}: decimals must be an integer from 0 to {MAX_DECIMALS}"
)
if spec.get("integer") and decimals:
raise ValueError(f"field {name}: an integer field has 0 decimals")
if kind == "estimate" and is_date_estimate(spec):
validate_date_range(name, spec)
elif kind == "estimate":
validate_number_range(name, spec)
def null_threshold(spec: dict[str, Any]) -> float:
"""A read field answers null when p_null is at or above this (default 0.5)."""
return float(spec.get("null_threshold", DEFAULT_NULL_THRESHOLD))
def _number(value: Any) -> float:
if isinstance(value, bool) or not isinstance(value, (int, float)):
raise TypeError(value)
return float(value)
def validate_number_range(
name: str, spec: dict[str, Any], kind: str = "estimate"
) -> None:
try:
low, high = _number(spec["min"]), _number(spec["max"])
except (KeyError, TypeError) as error:
raise ValueError(f"{kind} {name} needs numeric min and max") from error
if not (math.isfinite(low) and math.isfinite(high)) or low >= high:
raise ValueError(
f"{kind} {name} needs finite min < max, got {spec['min']!r} and {spec['max']!r}"
)
scale = spec.get("scale", "linear")
if scale not in ("linear", "log"):
raise ValueError(
f"{kind} {name} has an unknown scale {scale!r}; use 'linear' or 'log'"
)
if scale == "log" and low <= 0:
raise ValueError(
f"{kind} {name} uses scale 'log', which needs min > 0, got {spec['min']!r}"
)
if spec.get("step") is None:
return
try:
step = _number(spec["step"])
except TypeError as error:
raise ValueError(
f"{kind} {name}: step must be a positive number, got {spec['step']!r}"
) from error
if not math.isfinite(step) or step <= 0:
raise ValueError(
f"{kind} {name}: step must be a positive number, got {spec['step']!r}"
)
steps = (high - low) / step
if steps < 1:
raise ValueError(
f"{kind} {name}: step {spec['step']!r} is larger than the range {spec['min']!r} to {spec['max']!r}"
)
if steps > MAX_GRID_STEPS:
raise ValueError(
f"{kind} {name}: step {spec['step']!r} divides {spec['min']!r} to {spec['max']!r} into {steps:,.0f} steps; "
f"the limit is {MAX_GRID_STEPS:,}. Use a larger step or omit it"
)
def validate_reference_date(value: Any) -> None:
if value is None:
return
try:
datetime.date.fromisoformat(value)
except (TypeError, ValueError) as error:
raise ValueError(
f"reference_date must be an ISO date, got {value!r}"
) from error
def validate_date_range(
name: str, spec: dict[str, Any], kind: str = "estimate"
) -> None:
try:
low, high, _ = estimate_range(spec)
except (KeyError, TypeError, ValueError) as error:
raise ValueError(
f"date {kind} {name} needs ISO date min and max (YYYY-MM-DD)"
) from error
if not isinstance(spec.get("max"), str) or not low < high:
raise ValueError(
f"date {kind} {name} needs ISO date min < max, got {spec.get('min')!r} and {spec.get('max')!r}"
)
if spec.get("scale", "linear") != "linear":
raise ValueError(f"date {kind} {name} only takes scale 'linear'")
if date_granularity(spec) != "day" or spec.get("step") is None:
return
step = spec["step"]
if (
isinstance(step, bool)
or not isinstance(step, (int, float))
or not float(step).is_integer()
or step < 1
):
raise ValueError(
f"date {kind} {name}: step must be a whole number of days, 'month', or 'year', got {step!r}"
)
if (high - low) / step < 1:
raise ValueError(
f"date {kind} {name}: step {step!r} days is larger than the range"
)