clairic-flash / clairic_runtime /pdf_number_focus.py
areneau's picture
Add complete Clairic Transformers package
dc7d8aa verified
Raw History Blame Contribute Delete
2.91 kB
from __future__ import annotations
from collections.abc import Callable
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from .api import PredictRequest
from .backend import EncodedRequest
import re
from . import grounding, image_focus
VIEW_ID = "pdf-source-only-scalars-v1"
_PAGE_SCOPE = re.compile(
"\\b(?:page\\s+(?:[1-4]|one|two|three|four)|(?:first|second|third|fourth|1st|2nd|3rd|4th)\\s+page)\\b",
re.IGNORECASE,
)
_EXCLUDED_SCOPE = re.compile(
"\\b(?:not|ignore|exclude|excluding|except|without|rather than)\\b[^.;:\\n]*$",
re.IGNORECASE,
)
def _page_scoped(instructions: object) -> bool:
return isinstance(instructions, str) and any(
(
not _EXCLUDED_SCOPE.search(instructions[: match.start()])
for match in _PAGE_SCOPE.finditer(instructions)
)
)
def prepare(
request: PredictRequest,
encoded: EncodedRequest,
encode: Callable[..., EncodedRequest],
*,
enabled: bool,
) -> EncodedRequest:
if not enabled or not request.documents:
return encoded
(record, forecasts, tasks, typed, eos) = encoded
active = {name for (name, _, _) in tasks}
selected = {
name: spec
for (name, spec) in request.fields.items()
if name in active
and spec.get("type") == "number"
and (spec.get("source_only") is True)
and (not spec.get("knowledge"))
and (not grounding.state_only(spec))
and _page_scoped(spec.get("instructions", ""))
}
originals = record.get("_original_images", ())
pdf_pages = originals[len(request.images) :]
if (
not selected
or not pdf_pages
or (
not any(
(image_focus.focus_document(page) is not page for page in pdf_pages)
)
)
):
return encoded
child_request = request.model_copy(update={"fields": selected})
(child, _, replacements, child_typed, child_eos) = encode(
child_request,
identifier_view=True,
image_views=lambda: originals,
image_view_key=VIEW_ID,
)
if child_typed is not None or child_eos != eos:
raise ValueError("A literal PDF number view must retain the scalar AR contract")
by_name = {name: (name, ids, media) for (name, ids, media) in replacements}
if set(by_name) != set(selected):
raise ValueError(
"PDF number detail tasks differ from the selected public fields"
)
updated = [by_name.get(name, (name, ids, media)) for (name, ids, media) in tasks]
salts = dict(record.get("_ar_task_salts", {}))
for name in selected:
record["_ar_constraints"][name] = child["_ar_constraints"][name]
salts[name] = child["_cache_salt"]
record.update(
_ar_tasks=updated, _ar_task_salts=salts, _ar_prefix=[], _ar_group_prefixes=[]
)
return (record, forecasts, updated, typed, eos)