feat(evidence): validate canonical curated corpus
This commit is contained in:
@@ -0,0 +1,332 @@
|
||||
"""Typed, reviewable Evidence units stored in the workspace repository."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from pathlib import Path, PurePosixPath
|
||||
from typing import Literal
|
||||
|
||||
import sqlglot
|
||||
import yaml
|
||||
from pydantic import AnyHttpUrl, BaseModel, ConfigDict, Field, field_validator, model_validator
|
||||
from sqlglot import exp
|
||||
|
||||
from tht.evidence.contracts import validate_canonical_uri
|
||||
|
||||
|
||||
class StrictModel(BaseModel):
|
||||
"""Reject undeclared fields in the repository's canonical format."""
|
||||
|
||||
model_config = ConfigDict(extra="forbid")
|
||||
|
||||
|
||||
EvidenceKind = Literal[
|
||||
"glossary",
|
||||
"domain",
|
||||
"enum",
|
||||
"example",
|
||||
"mapping",
|
||||
"normalization",
|
||||
"formula",
|
||||
"reference",
|
||||
]
|
||||
EvidencePurpose = Literal[
|
||||
"disambiguation",
|
||||
"rewriting",
|
||||
"schema_linking",
|
||||
"sql_generation",
|
||||
]
|
||||
_IDENTIFIER = r"[A-Za-z_][A-Za-z0-9_$]*"
|
||||
_TABLE_IDENTIFIER = re.compile(rf"^{_IDENTIFIER}\.{_IDENTIFIER}$")
|
||||
_COLUMN_IDENTIFIER = re.compile(rf"^{_IDENTIFIER}\.{_IDENTIFIER}\.{_IDENTIFIER}$")
|
||||
MAX_CURATED_FILE_BYTES = 10 * 1024 * 1024
|
||||
|
||||
|
||||
class EvidenceScope(StrictModel):
|
||||
concepts: tuple[str, ...] = ()
|
||||
tables: tuple[str, ...] = ()
|
||||
columns: tuple[str, ...] = ()
|
||||
|
||||
@field_validator("tables")
|
||||
@classmethod
|
||||
def _validate_tables(cls, value: tuple[str, ...]) -> tuple[str, ...]:
|
||||
return _validate_identifiers(value, _TABLE_IDENTIFIER, "tables")
|
||||
|
||||
@field_validator("columns")
|
||||
@classmethod
|
||||
def _validate_columns(cls, value: tuple[str, ...]) -> tuple[str, ...]:
|
||||
return _validate_identifiers(value, _COLUMN_IDENTIFIER, "columns")
|
||||
|
||||
|
||||
class EvidenceProvenance(StrictModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True)
|
||||
|
||||
source_file: str
|
||||
source_sha256: str
|
||||
supporting_excerpts: tuple[str, ...]
|
||||
|
||||
@field_validator("source_file")
|
||||
@classmethod
|
||||
def _validate_source_file(cls, value: str) -> str:
|
||||
return validate_source_file(value)
|
||||
|
||||
@field_validator("source_sha256")
|
||||
@classmethod
|
||||
def _validate_sha256(cls, value: str) -> str:
|
||||
if not re.fullmatch(r"sha256:[0-9a-f]{64}", value):
|
||||
raise ValueError("source_sha256 must be a sha256 digest")
|
||||
return value
|
||||
|
||||
@field_validator("supporting_excerpts")
|
||||
@classmethod
|
||||
def _validate_excerpts(cls, value: tuple[str, ...]) -> tuple[str, ...]:
|
||||
if not 1 <= len(value) <= 5:
|
||||
raise ValueError("supporting_excerpts must contain one to five items")
|
||||
if any(not excerpt.strip() or len(excerpt) > 1000 for excerpt in value):
|
||||
raise ValueError("supporting excerpts must be nonempty and at most 1000 characters")
|
||||
return value
|
||||
|
||||
|
||||
class ReviewItem(StrictModel):
|
||||
code: str
|
||||
message: str
|
||||
field: str | None = None
|
||||
|
||||
|
||||
class FormulaPayload(StrictModel):
|
||||
concept: str
|
||||
columns: tuple[str, ...]
|
||||
sql: str
|
||||
|
||||
@field_validator("columns")
|
||||
@classmethod
|
||||
def _validate_columns(cls, value: tuple[str, ...]) -> tuple[str, ...]:
|
||||
return _validate_identifiers(value, _COLUMN_IDENTIFIER, "columns")
|
||||
|
||||
@field_validator("sql")
|
||||
@classmethod
|
||||
def _validate_expression(cls, value: str) -> str:
|
||||
try:
|
||||
statements = [statement for statement in sqlglot.parse(value, read="postgres") if statement]
|
||||
except sqlglot.errors.ParseError as error:
|
||||
raise ValueError("formula.sql must be valid PostgreSQL") from error
|
||||
if len(statements) != 1:
|
||||
raise ValueError("formula.sql must contain exactly one expression")
|
||||
expression = statements[0]
|
||||
if expression.find(exp.Select) is not None or expression.find(exp.With) is not None:
|
||||
raise ValueError("formula.sql must not contain a query")
|
||||
if any(
|
||||
expression.find(statement_type) is not None
|
||||
for statement_type in (
|
||||
exp.Insert,
|
||||
exp.Update,
|
||||
exp.Delete,
|
||||
exp.Create,
|
||||
exp.Drop,
|
||||
exp.Alter,
|
||||
exp.Merge,
|
||||
exp.TruncateTable,
|
||||
exp.Grant,
|
||||
exp.Revoke,
|
||||
exp.Command,
|
||||
exp.Values,
|
||||
exp.Set,
|
||||
exp.Table,
|
||||
)
|
||||
):
|
||||
raise ValueError("formula.sql must not contain DDL or DML")
|
||||
return value
|
||||
|
||||
|
||||
class ReferencePayload(StrictModel):
|
||||
url: AnyHttpUrl
|
||||
label: str
|
||||
description: str
|
||||
|
||||
@field_validator("url")
|
||||
@classmethod
|
||||
def _reject_credentials(cls, value: AnyHttpUrl) -> AnyHttpUrl:
|
||||
validate_canonical_uri(str(value))
|
||||
return value
|
||||
|
||||
|
||||
class GlossaryPayload(StrictModel):
|
||||
definition: str
|
||||
synonyms: tuple[str, ...] = ()
|
||||
variants: tuple[str, ...] = ()
|
||||
|
||||
|
||||
class DomainPayload(StrictModel):
|
||||
rule: str
|
||||
|
||||
|
||||
class EnumPayload(StrictModel):
|
||||
column: str
|
||||
values: dict[str, str]
|
||||
|
||||
@field_validator("column")
|
||||
@classmethod
|
||||
def _validate_column(cls, value: str) -> str:
|
||||
_validate_identifiers((value,), _COLUMN_IDENTIFIER, "column")
|
||||
return value
|
||||
|
||||
|
||||
class ExamplePayload(StrictModel):
|
||||
question: str
|
||||
interpretation: str
|
||||
|
||||
|
||||
class MappingPayload(StrictModel):
|
||||
concept: str
|
||||
tables: tuple[str, ...]
|
||||
columns: tuple[str, ...]
|
||||
|
||||
@field_validator("tables")
|
||||
@classmethod
|
||||
def _validate_tables(cls, value: tuple[str, ...]) -> tuple[str, ...]:
|
||||
return _validate_identifiers(value, _TABLE_IDENTIFIER, "tables")
|
||||
|
||||
@field_validator("columns")
|
||||
@classmethod
|
||||
def _validate_columns(cls, value: tuple[str, ...]) -> tuple[str, ...]:
|
||||
return _validate_identifiers(value, _COLUMN_IDENTIFIER, "columns")
|
||||
|
||||
|
||||
class NormalizationPayload(StrictModel):
|
||||
input: str
|
||||
output: str
|
||||
rule: str
|
||||
|
||||
|
||||
EvidencePayload = (
|
||||
GlossaryPayload
|
||||
| DomainPayload
|
||||
| EnumPayload
|
||||
| ExamplePayload
|
||||
| MappingPayload
|
||||
| NormalizationPayload
|
||||
| FormulaPayload
|
||||
| ReferencePayload
|
||||
)
|
||||
|
||||
|
||||
_PAYLOAD_TYPE_BY_KIND = {
|
||||
"glossary": GlossaryPayload,
|
||||
"domain": DomainPayload,
|
||||
"enum": EnumPayload,
|
||||
"example": ExamplePayload,
|
||||
"mapping": MappingPayload,
|
||||
"normalization": NormalizationPayload,
|
||||
"formula": FormulaPayload,
|
||||
"reference": ReferencePayload,
|
||||
}
|
||||
_EVIDENCE_ID = re.compile(r"^evidence:[a-z0-9]+(?:-[a-z0-9]+)*$")
|
||||
|
||||
|
||||
class CuratedEvidence(StrictModel):
|
||||
schema_version: Literal[1]
|
||||
id: str
|
||||
title: str
|
||||
kind: EvidenceKind
|
||||
purposes: tuple[EvidencePurpose, ...]
|
||||
applies_to: EvidenceScope = Field(default_factory=EvidenceScope)
|
||||
language: str
|
||||
provenance: EvidenceProvenance
|
||||
review_items: tuple[ReviewItem, ...] = ()
|
||||
payload: EvidencePayload
|
||||
|
||||
@model_validator(mode="after")
|
||||
def _validate_kind_payload(self) -> CuratedEvidence:
|
||||
if not is_evidence_id(self.id):
|
||||
raise ValueError("id must use the evidence:<slug> form")
|
||||
expected = _PAYLOAD_TYPE_BY_KIND.get(self.kind)
|
||||
if expected is not None and not isinstance(self.payload, expected):
|
||||
raise ValueError(f"{self.kind} requires its typed payload")
|
||||
return self
|
||||
|
||||
|
||||
def parse_curated_markdown(text: str, *, path: Path | None = None) -> CuratedEvidence:
|
||||
"""Parse the canonical frontmatter representation of one Curated Evidence unit."""
|
||||
if not text.startswith("---\n"):
|
||||
raise ValueError("curated evidence requires YAML frontmatter")
|
||||
try:
|
||||
_, frontmatter, body = text.split("---\n", 2)
|
||||
except ValueError as error:
|
||||
raise ValueError("curated evidence frontmatter is malformed") from error
|
||||
raw = yaml.safe_load(frontmatter)
|
||||
try:
|
||||
data = dict(raw)
|
||||
except (TypeError, ValueError) as error:
|
||||
raise ValueError("curated evidence frontmatter must be a mapping") from error
|
||||
if body.strip():
|
||||
raise ValueError("curated evidence must not contain an ignored body")
|
||||
kind = data.get("kind")
|
||||
if "payload" not in data and kind in _PAYLOAD_TYPE_BY_KIND:
|
||||
data["payload"] = data.pop(kind, None)
|
||||
evidence = CuratedEvidence.model_validate(data)
|
||||
if path is not None:
|
||||
_validate_kind_directory(path, evidence.kind)
|
||||
return evidence
|
||||
|
||||
|
||||
def dump_curated_markdown(value: CuratedEvidence) -> str:
|
||||
"""Render canonical frontmatter with a human-readable kind-specific payload key."""
|
||||
data = value.model_dump(mode="json", exclude={"payload"})
|
||||
data[value.kind] = value.payload.model_dump(mode="json")
|
||||
frontmatter = yaml.safe_dump(data, allow_unicode=True, sort_keys=False)
|
||||
return f"---\n{frontmatter}---\n"
|
||||
|
||||
|
||||
def load_curated_tree(root: Path) -> list[CuratedEvidence]:
|
||||
"""Load canonical Evidence units in stable path order from a curated root."""
|
||||
if not root.is_dir():
|
||||
return []
|
||||
documents: list[CuratedEvidence] = []
|
||||
for path in sorted(root.rglob("*.md")):
|
||||
if path.name.upper().startswith("README"):
|
||||
continue
|
||||
if path.stat().st_size > MAX_CURATED_FILE_BYTES:
|
||||
raise ValueError("curated evidence exceeds the size limit")
|
||||
try:
|
||||
text = path.read_text(encoding="utf-8")
|
||||
except UnicodeDecodeError as error:
|
||||
raise ValueError("curated evidence must be UTF-8") from error
|
||||
documents.append(parse_curated_markdown(text, path=path))
|
||||
return documents
|
||||
|
||||
|
||||
def _validate_kind_directory(path: Path, kind: EvidenceKind) -> None:
|
||||
parts = path.parts
|
||||
try:
|
||||
curated_index = parts.index("curated")
|
||||
except ValueError:
|
||||
return
|
||||
if len(parts) <= curated_index + 1 or parts[curated_index + 1] != kind:
|
||||
raise ValueError("curated evidence kind must match its directory")
|
||||
|
||||
|
||||
def _validate_identifiers(
|
||||
values: tuple[str, ...], pattern: re.Pattern[str], field: str,
|
||||
) -> tuple[str, ...]:
|
||||
if any(pattern.fullmatch(value) is None for value in values):
|
||||
raise ValueError(f"{field} must use canonical schema identifiers")
|
||||
return values
|
||||
|
||||
|
||||
def validate_source_file(value: str) -> str:
|
||||
"""Validate a repository-relative, credential-free Source Evidence path."""
|
||||
path = PurePosixPath(value)
|
||||
if (
|
||||
path.is_absolute()
|
||||
or ".." in path.parts
|
||||
or not path.parts
|
||||
or path.parts[0] != "source"
|
||||
or not value.endswith((".md", ".txt", ".sql.md"))
|
||||
):
|
||||
raise ValueError("source_file must be a supported path below source/")
|
||||
return value
|
||||
|
||||
|
||||
def is_evidence_id(value: str) -> bool:
|
||||
"""Whether a value uses the stable public Evidence identifier format."""
|
||||
return _EVIDENCE_ID.fullmatch(value) is not None
|
||||
Reference in New Issue
Block a user