vibe.review.models¶

Data models for VIBE Review.

This module defines both: - SQLAlchemy models for database persistence - Dataclasses for in-memory representation and API boundaries

The design separates concerns: - SQLAlchemy models handle persistence and querying - Dataclasses provide clean interfaces for business logic

Base ¶

SQLAlchemy declarative base for all models.

PartType ¶

Type of document part.

Unified enum covering both structural types (heading, paragraph, etc.) and semantic/contract-specific types (clause, section, article, etc.).

ClassificationResult ¶

Possible classification outcomes for a requirement review.

DocumentStatus ¶

Document processing status.

Requirement ¶

A single compliance requirement to be checked against a document.

Attributes:

id (str) –

Unique identifier (e.g., "D2-1", "GDPR-7")
label (str) –

Human-readable short label
description (str) –

Full requirement description
reference_text (str | None) –

Original regulatory text this derives from
reference_source (str | None) –

Source document/article reference
group_id (str | None) –

The group this requirement belongs to (single group, set from nesting)
help (str | None) –

Evaluation guidance for reviewers (explains how to assess)
embedding (list[float] | None) –

Pre-computed embedding vector (if available)

Note: Requirement applicability is determined by template probing, not by conditions in config. The template's Jinja conditionals (e.g., {% if critical_service %}) control which requirements apply - "template is truth".

build_query_text ¶

build_query_text() -> str

Build query text for search and reranking.

Concatenates label, description, reference (with source if available), and help text. Used by both hybrid search and cross-encoder reranking.

to_dict ¶

to_dict() -> dict[str, Any]

Serialize requirement to dictionary, excluding embedding.

from_dict ¶

from_dict(data: dict[str, Any]) -> Requirement

Reconstruct requirement from dictionary; embedding will be None.

RequirementGroup ¶

A logical grouping of related requirements.

Groups are used for: - Organizing the review UI - Generating section-based reports - Conditional applicability (e.g., "critical services" group)

ReviewClassification ¶

A classification decision for a requirement against a document.

Captures both AI-suggested and human-verified classifications.

ReviewProgress ¶

Progress tracking for a review session.

ClassificationStats ¶

Classification statistics for a review session.

RequirementAccuracyStats ¶

Accuracy statistics for a single requirement.

AccuracyStats ¶

AI classification accuracy statistics.

TemplateInfo ¶

Basic template metadata.

AssessmentNavigation ¶

Navigation state for the assessment stream.

MatchedPart ¶

A document part matched to a requirement.

QuestionReview ¶

Review decision for a template question (parallel to ReviewClassification).

SessionListRow ¶

Row data for the review session list (index.html).

TemplateSessionRow ¶

Row data for template-specific session list (template_sessions.html).

RequirementCacheModel ¶

Cached requirement with pre-computed embedding.

Requirements are loaded from template config.yml but cached in DB with embeddings to avoid recomputation.

to_dataclass ¶

to_dataclass() -> Requirement

Convert to Requirement dataclass, including embedding vector.

DocumentModel ¶

A document submitted for compliance review.

display_name ¶

display_name: str

Return label if set, otherwise filename.

DocumentPartModel ¶

A segment of a document with embedding for retrieval.

This is the canonical persistence model for document parts. SemanticUnit instances are converted to this model via to_db_fields() for storage.

ReviewSessionModel ¶

A review session tracking progress through requirement checks.

RequirementReviewModel ¶

Individual requirement classification within a review session.

to_dataclass ¶

to_dataclass() -> ReviewClassification

Convert to ReviewClassification dataclass with all fields.

QuestionReviewModel ¶

Individual question review within a review session.

Tracks AI suggestions vs human answers for template questions, parallel to RequirementReviewModel for requirements.

effective_answer ¶

effective_answer: object

Return human override if verified, otherwise AI-suggested answer.

to_dataclass ¶

to_dataclass() -> QuestionReview

Convert to QuestionReview dataclass with all fields.

ExampleModel ¶

Few-shot examples for requirement classification.

Examples are used to guide LLM classification with known-good cases. Can optionally track provenance back to the source review and document.

ReferenceSourceModel ¶

A regulatory source document (e.g., DORA, EBA Guidelines).

Stores metadata about the source and links to its parts.

ReferencePartModel ¶

A searchable part of a reference source document.

Parts are individual articles, recitals, or guideline paragraphs that can be retrieved and cited in compliance reviews.

utc_now ¶

utc_now() -> datetime

Return the current UTC time.