anyformat 1.2.0__tar.gz → 1.3.0__tar.gz

{anyformat-1.2.0 → anyformat-1.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: anyformat
-Version: 1.2.0
+Version: 1.3.0
 Summary: Python SDK for the AnyFormat API
 Author-email: Anyformat <info@anyformat.ai>
 Requires-Python: >=3.10

{anyformat-1.2.0 → anyformat-1.3.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "anyformat"
-version = "1.2.0"
+version = "1.3.0"
 description = "Python SDK for the AnyFormat API"
 readme = "README.md"
 authors = [

{anyformat-1.2.0 → anyformat-1.3.0}/src/anyformat/sdk/_viz.py

@@ -15,6 +15,8 @@ from anyformat.sdk._source import SourceDocument
 
 if TYPE_CHECKING:
     from PIL.Image import Image
+    from PIL.ImageDraw import ImageDraw
+    from PIL.ImageFont import FreeTypeFont, ImageFont
 
 _EMERALD = (74, 222, 128)  # #4ADE80  high  (>=80)
 _AMBER = (251, 191, 36)  # #FBBF24  medium (50-79)
@@ -110,7 +112,9 @@ def _draw_on_page(page: "Image", boxes: list[Box]) -> "Image":
     return Image.alpha_composite(base, overlay).convert("RGB")
 
 
-def _draw_label(draw, x: float, y: float, text: str, color: tuple[int, int, int], font) -> None:
+def _draw_label(
+    draw: "ImageDraw", x: float, y: float, text: str, color: tuple[int, int, int], font: "FreeTypeFont | ImageFont"
+) -> None:
     left, top, right, bottom = draw.textbbox((0, 0), text, font=font)
     pad = 2
     tag_w = (right - left) + 2 * pad
@@ -119,7 +123,7 @@ def _draw_label(draw, x: float, y: float, text: str, color: tuple[int, int, int]
     draw.text((x + pad, y + pad), text, fill=(255, 255, 255, 255), font=font)
 
 
-def _load_font(size: int):
+def _load_font(size: int) -> "FreeTypeFont | ImageFont":
     from PIL import ImageFont
 
     return ImageFont.load_default(size=size)

{anyformat-1.2.0 → anyformat-1.3.0}/src/anyformat/sdk/cli.py

@@ -179,7 +179,7 @@ class OnConflictMode(str, Enum):
 
 
 @app.callback(invoke_without_command=True)
-def main(
+def root(
     ctx: typer.Context,
     version_flag: Annotated[
         bool,

{anyformat-1.2.0 → anyformat-1.3.0}/src/anyformat/workflow/fields.py

@@ -71,7 +71,7 @@ class _Base(BaseModel):
 
     @model_validator(mode="before")
     @classmethod
-    def _inject_data_type(cls, data):
+    def _inject_data_type(cls, data: object) -> object:
         # Direct Python construction (``StringField(name=..., description=...)``)
         # should not require repeating the discriminator literal — the subclass
         # already pins it. Inject it from the ``Literal[…]`` annotation when

{anyformat-1.2.0 → anyformat-1.3.0}/src/anyformat/workflow/results.py

@@ -5,9 +5,16 @@ This is the canonical response shape — every cookbook recipe and the OpenAPI s
 mirror it.
 """
 
-from typing import Any, Generic, Optional, TypeVar, Union
+from typing import Any, Generic, Literal, Optional, TypeVar, Union
 
-from pydantic import BaseModel, Field, field_validator, model_serializer, model_validator
+from pydantic import (
+    BaseModel,
+    Field,
+    SerializerFunctionWrapHandler,
+    field_validator,
+    model_serializer,
+    model_validator,
+)
 
 # Generic over the extracted ``value`` / ``value_override`` so SDK consumers
 # can parameterise ``ExtractedField[float]`` etc. and stop seeing ``Any``. The
@@ -30,6 +37,30 @@ class Evidence(BaseModel):
     page_number: int = Field(description="1-indexed page number where the snippet was found.")
 
 
+class ConsolidationCandidate(BaseModel):
+    """One part of a long document's answer for a field whose parts disagreed."""
+
+    value: Optional[Any] = Field(default=None, description="The value this part of the document gave.")
+    value_unit: Optional[str] = Field(default=None, description="The unit of `value`, when it has one.")
+    evidence: list[Evidence] = Field(
+        default_factory=list, description="Source-text snippets this part cited. Empty when it cited none."
+    )
+
+    @field_validator("evidence", mode="before")
+    @classmethod
+    def _coerce_null_evidence(cls, value: Any) -> Any:
+        return [] if value is None else value
+
+
+class Consolidation(BaseModel):
+    """Why a field of a long document has no value: its parts gave answers that could not be combined."""
+
+    status: Literal["unresolved"] = Field(description="Always `unresolved`: no value could be chosen.")
+    candidates: list[ConsolidationCandidate] = Field(
+        description="Each part's answer, to compare against the document before entering the value."
+    )
+
+
 class ExtractedField(BaseModel, Generic[ValueT]):
     """One extracted field's value, confidence, and supporting evidence."""
 
@@ -58,6 +89,20 @@ class ExtractedField(BaseModel, Generic[ValueT]):
         default_factory=list,
         description="Source-text snippets the model used to derive this value.",
     )
+    value_source: Optional[Literal["document", "lookup"]] = Field(
+        default=None,
+        description="Where a smart-lookup field's value came from. Omitted for other fields.",
+    )
+    lookup_check: Optional[Literal["confirmed", "conflict"]] = Field(
+        default=None,
+        description="Whether the lookup `confirmed` the document value or was in `conflict` with it (the document "
+        "value is kept). Omitted when no lookup compared the two.",
+    )
+    consolidation: Optional[Consolidation] = Field(
+        default=None,
+        description="Set when a long document's parts gave answers that could not be combined; `value` is then "
+        "`null`. Omitted otherwise.",
+    )
 
     @field_validator("evidence", mode="before")
     @classmethod
@@ -66,6 +111,15 @@ class ExtractedField(BaseModel, Generic[ValueT]):
         # when a field has no extracted value; collapse that to an empty list.
         return [] if value is None else value
 
+    @model_serializer(mode="wrap")
+    def _drop_unset_provenance(self, handler: SerializerFunctionWrapHandler):  # noqa: ANN202 — pydantic reads a return annotation as the serialization schema
+        # Only lookup or unresolved fields carry these; every other field keeps its shape.
+        data = handler(self)
+        for key in ("value_source", "lookup_check", "consolidation"):
+            if data.get(key) is None:
+                data.pop(key, None)
+        return data
+
 
 # A scalar extracted field, or a list of object-field rows — each row is a dict
 # from cell name to ExtractedField (the shape produced by object-type extraction).
@@ -266,11 +320,9 @@ class ClassificationResult(BaseModel):
 
 
 class Validation(BaseModel):
-    """A configured rule verdict or an automatic extraction review warning."""
+    """A configured rule verdict."""
 
-    rule_id: str = Field(
-        description="The configured rule ID, or an `extract.consolidation:` ID for an automatic review warning."
-    )
+    rule_id: str = Field(description="The configured rule ID.")
     description: str = Field(
         default="",
         description="The rule's natural-language description; empty for a deterministic rule.",
@@ -313,11 +365,7 @@ class Extraction(BaseModel):
     )
     validations: list[Validation] = Field(
         default_factory=list,
-        description=(
-            "Configured validation verdicts and automatic extraction review warnings. "
-            "Unresolved chunk fields have a null value and an inconclusive warning with candidate values and evidence, "
-            "even without a validate node."
-        ),
+        description="Verdicts of the workflow's validate nodes.",
     )
 
 
@@ -484,7 +532,7 @@ class ResultsResponseV3(ResultsResponse):
     model_config = {"json_schema_extra": _v3_schema_transform}
 
     @model_serializer(mode="wrap")
-    def _drop_legacy_collection_id(self, handler):
+    def _drop_legacy_collection_id(self, handler: SerializerFunctionWrapHandler):  # noqa: ANN202 — pydantic reads a return annotation as the serialization schema
         data = handler(self)
         data.pop("collection_id", None)
         return data

{anyformat-1.2.0 → anyformat-1.3.0}/src/anyformat/workflow/topology.py

@@ -618,6 +618,35 @@ class NoParseAsBranchTarget:
         ]
 
 
+class NoParseUnderSplitter:
+    """Reject a Parse reachable from a Splitter, however deep. A Parse inside a
+    partition re-parses every file of the packet once per partition, not the slice."""
+
+    name = "no_parse_under_splitter"
+
+    def check(self, graph: "WorkflowDefinition") -> list[Violation]:
+        successors: dict[str, list[str]] = {}
+        for edge in graph.edges:
+            successors.setdefault(edge.source, []).append(edge.target)
+        frontier = [n.id for n in graph.nodes if isinstance(n, SplitterNode)]
+        downstream_of_splitter: set[str] = set()
+        while frontier:
+            for target in successors.get(frontier.pop(), []):
+                if target not in downstream_of_splitter:
+                    downstream_of_splitter.add(target)
+                    frontier.append(target)
+        return [
+            Violation(
+                rule=self.name,
+                message=f"parse node {node.id!r} sits downstream of a splitter; a parse "
+                "cannot run inside a partition.",
+                node_ids=[node.id],
+            )
+            for node in graph.nodes
+            if isinstance(node, ParseNode) and node.id in downstream_of_splitter
+        ]
+
+
 class EveryExtractNodeHasFields:
     """Belt-and-braces redundancy with ``ExtractionSchema.fields min_length=1``:
     pydantic catches the empty-fields case first at the wire boundary, but
@@ -839,6 +868,7 @@ RULES: list[TopologyRule] = [
     EveryExtractNodeHasFields(),
     EveryIfElseHasTrueBranch(),
     NoParseAsBranchTarget(),
+    NoParseUnderSplitter(),
     NoCodeSandboxTemplate(),
     CodeSandboxTemplateFollowsExtract(),
     ValidationRuleIdReferencesExist(),
@@ -894,6 +924,7 @@ __all__ = [
     "ExactlyOneParseNode",
     "NoCycles",
     "NoParseAsBranchTarget",
+    "NoParseUnderSplitter",
     "NonBranchingNodeHasAtMostOneOutgoing",
     "RULES",
     "SingleUpstreamNodeHasOneIncoming",