feat(evidence): build semantic fragments from typed units
This commit is contained in:
@@ -2,6 +2,7 @@ import hashlib
|
||||
|
||||
import pytest
|
||||
|
||||
from tht.evidence.canonical import CuratedEvidence
|
||||
from tht.evidence.corpus.chunk import ChunkPolicy, chunk
|
||||
from tht.evidence.corpus.models import CanonicalDocument, CorpusManifest
|
||||
|
||||
@@ -33,6 +34,50 @@ def other_document(content: str) -> CanonicalDocument:
|
||||
)
|
||||
|
||||
|
||||
def curated_formula_document(*, sql: str = "CASE WHEN age < 18 THEN 'pediatric' END") -> CanonicalDocument:
|
||||
evidence = CuratedEvidence.model_validate(
|
||||
{
|
||||
"schema_version": 1,
|
||||
"id": "evidence:fascia-pediatrica",
|
||||
"title": "Fascia pediatrica",
|
||||
"kind": "formula",
|
||||
"purposes": ["sql_generation", "schema_linking"],
|
||||
"applies_to": {
|
||||
"concepts": ["fascia pediatrica"],
|
||||
"tables": ["clinical.patient"],
|
||||
"columns": ["clinical.patient.birth_date"],
|
||||
},
|
||||
"language": "it",
|
||||
"provenance": {
|
||||
"source_file": "source/paziente.md",
|
||||
"source_sha256": "sha256:" + "a" * 64,
|
||||
"supporting_excerpts": ["I pazienti pediatrici hanno età inferiore a 18 anni."],
|
||||
},
|
||||
"review_items": [],
|
||||
"payload": {
|
||||
"concept": "fascia pediatrica",
|
||||
"columns": ["clinical.patient.birth_date"],
|
||||
"sql": sql,
|
||||
},
|
||||
}
|
||||
)
|
||||
return curated_document(evidence)
|
||||
|
||||
|
||||
def curated_document(evidence: CuratedEvidence) -> CanonicalDocument:
|
||||
content = "canonical curated Evidence"
|
||||
return CanonicalDocument(
|
||||
document_id="doc:" + evidence.id.removeprefix("evidence:"),
|
||||
source_id="curated:" + evidence.id.removeprefix("evidence:"),
|
||||
source_uri=f"file:///safe/curated/{evidence.kind}/{evidence.id.removeprefix('evidence:')}.md",
|
||||
source_fingerprint="sha256:" + "b" * 64,
|
||||
content_hash="sha256:" + hashlib.sha256(content.encode()).hexdigest(),
|
||||
title=evidence.title,
|
||||
content=content,
|
||||
media_type="text/markdown",
|
||||
pipeline_version="pipe:v1",
|
||||
metadata={"curated_evidence": evidence.model_dump(mode="json")},
|
||||
)
|
||||
def test_chunk_ids_are_stable_for_same_content_and_repeat_runs():
|
||||
policy = ChunkPolicy(version="paragraph:v1", max_chars=8)
|
||||
first = chunk(document("A\n\nB"), policy)
|
||||
@@ -116,3 +161,128 @@ def test_empty_document_has_no_chunks_and_invalid_policy_is_rejected():
|
||||
assert chunk(document(""), ChunkPolicy(version="v1", max_chars=4)) == []
|
||||
with pytest.raises(ValueError):
|
||||
ChunkPolicy(version="v1", max_chars=0)
|
||||
|
||||
|
||||
def test_typed_formula_is_rendered_as_one_traceable_semantic_fragment():
|
||||
fragments = chunk(curated_formula_document(), ChunkPolicy(version="semantic:v1", max_chars=4000))
|
||||
|
||||
assert len(fragments) == 1
|
||||
fragment = fragments[0]
|
||||
assert "Formula: Fascia pediatrica" in fragment.content
|
||||
assert "Scopi: sql_generation, schema_linking" in fragment.content
|
||||
assert "Concetto: fascia pediatrica" in fragment.content
|
||||
assert "Colonne: clinical.patient.birth_date" in fragment.content
|
||||
assert "SQL: CASE WHEN age < 18 THEN 'pediatric' END" in fragment.content
|
||||
assert "Provenienza: source/paziente.md" in fragment.content
|
||||
assert fragment.metadata["evidence_id"] == "evidence:fascia-pediatrica"
|
||||
assert fragment.metadata["evidence_kind"] == "formula"
|
||||
assert fragment.metadata["purposes"] == ["sql_generation", "schema_linking"]
|
||||
assert fragment.metadata["scope"] == {
|
||||
"concepts": ["fascia pediatrica"],
|
||||
"tables": ["clinical.patient"],
|
||||
"columns": ["clinical.patient.birth_date"],
|
||||
}
|
||||
assert fragment.metadata["language"] == "it"
|
||||
assert fragment.metadata["provenance"]["source_file"] == "source/paziente.md"
|
||||
|
||||
|
||||
def test_oversized_typed_atomic_content_fails_instead_of_being_split():
|
||||
document = curated_formula_document(sql="CASE WHEN age < 18 THEN " + "x" * 200 + " END")
|
||||
|
||||
with pytest.raises(ValueError, match="atomic_content_too_large") as caught:
|
||||
chunk(document, ChunkPolicy(version="semantic:v1", max_chars=120))
|
||||
|
||||
assert caught.value.review_item.code == "atomic_content_too_large"
|
||||
assert caught.value.review_item.field == "formula.sql"
|
||||
|
||||
|
||||
def test_enum_value_meaning_pairs_are_atomic_and_fragment_ids_are_deterministic():
|
||||
payload = curated_formula_document().metadata["curated_evidence"]
|
||||
evidence = CuratedEvidence.model_validate({
|
||||
**payload,
|
||||
"id": "evidence:stato-ricovero",
|
||||
"title": "Stato ricovero",
|
||||
"kind": "enum",
|
||||
"payload": {
|
||||
"column": "clinical.admission.status",
|
||||
"values": {"A": "Attivo", "D": "Dimesso"},
|
||||
},
|
||||
})
|
||||
document = curated_document(evidence)
|
||||
|
||||
first = chunk(document, ChunkPolicy(version="semantic:v1", max_chars=4000))
|
||||
second = chunk(document, ChunkPolicy(version="semantic:v1", max_chars=4000))
|
||||
|
||||
assert first == second
|
||||
assert [fragment.ordinal for fragment in first] == [0, 1]
|
||||
assert "Valore: A\nSignificato: Attivo" in first[0].content
|
||||
assert "Valore: D\nSignificato: Dimesso" in first[1].content
|
||||
|
||||
|
||||
def test_italian_evidence_uses_an_italian_kind_heading():
|
||||
base = curated_formula_document().metadata["curated_evidence"]
|
||||
evidence = CuratedEvidence.model_validate({
|
||||
**base,
|
||||
"id": "evidence:regola-ricovero",
|
||||
"title": "Regola ricovero",
|
||||
"kind": "domain",
|
||||
"payload": {"rule": "Il ricovero richiede una data di ammissione."},
|
||||
})
|
||||
|
||||
fragment = chunk(curated_document(evidence), ChunkPolicy(version="semantic:v1", max_chars=4000))[0]
|
||||
|
||||
assert fragment.content.startswith("Dominio: Regola ricovero\n")
|
||||
|
||||
|
||||
def test_domain_rule_remains_atomic_across_blank_paragraphs():
|
||||
base = curated_formula_document().metadata["curated_evidence"]
|
||||
evidence = CuratedEvidence.model_validate({
|
||||
**base,
|
||||
"id": "evidence:regole-ricovero",
|
||||
"title": "Regole ricovero",
|
||||
"kind": "domain",
|
||||
"payload": {"rule": "La data di ammissione è obbligatoria.\n\nLa data di dimissione segue l'ammissione."},
|
||||
})
|
||||
|
||||
fragments = chunk(curated_document(evidence), ChunkPolicy(version="semantic:v1", max_chars=4000))
|
||||
|
||||
assert len(fragments) == 1
|
||||
assert "Regola: La data di ammissione è obbligatoria.\n\nLa data di dimissione segue l'ammissione." in fragments[0].content
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("kind", "payload", "field"),
|
||||
[
|
||||
("domain", {"rule": "x" * 200}, "domain.rule"),
|
||||
(
|
||||
"mapping",
|
||||
{
|
||||
"concept": "ricovero",
|
||||
"tables": ["clinical.admission"],
|
||||
"columns": ["clinical.admission.status"],
|
||||
},
|
||||
"mapping",
|
||||
),
|
||||
(
|
||||
"reference",
|
||||
{"url": "https://example.test/guide", "label": "Guida", "description": "x" * 200},
|
||||
"reference.url",
|
||||
),
|
||||
("normalization", {"input": "a", "output": "b", "rule": "x" * 200}, "normalization.rule"),
|
||||
("glossary", {"definition": "x" * 200}, "glossary.definition"),
|
||||
("example", {"question": "x" * 200, "interpretation": "attesa"}, "example"),
|
||||
],
|
||||
)
|
||||
def test_other_typed_atomic_content_blocks_with_a_stable_review_item(kind, payload, field):
|
||||
base = curated_formula_document().metadata["curated_evidence"]
|
||||
evidence = CuratedEvidence.model_validate({
|
||||
**base,
|
||||
"id": f"evidence:{kind}-test",
|
||||
"kind": kind,
|
||||
"payload": payload,
|
||||
})
|
||||
|
||||
with pytest.raises(ValueError, match="atomic_content_too_large") as caught:
|
||||
chunk(curated_document(evidence), ChunkPolicy(version="semantic:v1", max_chars=120))
|
||||
|
||||
assert caught.value.review_item.field == field
|
||||
|
||||
Reference in New Issue
Block a user