Files
syndarix/backend/app/models/memory/fact.py
Felipe Cardoso bd988f76b0 fix(memory): address review findings from Issue #88
Fixes based on multi-agent review:

Model Improvements:
- Remove duplicate index ix_procedures_agent_type (already indexed via Column)
- Fix postgresql_where to use text() instead of string literal in Fact model
- Add thread-safety to Procedure.success_rate property (snapshot values)

Data Integrity Constraints:
- Add CheckConstraint for Episode: importance_score 0-1, duration >= 0, tokens >= 0
- Add CheckConstraint for Fact: confidence 0-1
- Add CheckConstraint for Procedure: success_count >= 0, failure_count >= 0

Migration Updates:
- Add check constraints creation in upgrade()
- Add check constraints removal in downgrade()

Note: SQLAlchemy Column default=list is correct (callable factory pattern)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-05 01:54:51 +01:00

111 lines
3.2 KiB
Python

# app/models/memory/fact.py
"""
Fact database model.
Stores semantic memories - learned facts in subject-predicate-object
triple format with confidence scores and source tracking.
"""
from sqlalchemy import (
CheckConstraint,
Column,
DateTime,
Float,
ForeignKey,
Index,
Integer,
String,
Text,
text,
)
from sqlalchemy.dialects.postgresql import (
ARRAY,
UUID as PGUUID,
)
from sqlalchemy.orm import relationship
from app.models.base import Base, TimestampMixin, UUIDMixin
# Import pgvector type
try:
from pgvector.sqlalchemy import Vector # type: ignore[import-not-found]
except ImportError:
Vector = None
class Fact(Base, UUIDMixin, TimestampMixin):
"""
Semantic memory model.
Stores learned facts as subject-predicate-object triples:
- "FastAPI" - "uses" - "Starlette framework"
- "Project Alpha" - "requires" - "OAuth authentication"
Facts have confidence scores that decay over time and can be
reinforced when the same fact is learned again.
"""
__tablename__ = "facts"
# Scoping: project_id is NULL for global facts
project_id = Column(
PGUUID(as_uuid=True),
ForeignKey("projects.id", ondelete="CASCADE"),
nullable=True,
index=True,
)
# Triple format
subject = Column(String(500), nullable=False, index=True)
predicate = Column(String(255), nullable=False, index=True)
object = Column(Text, nullable=False)
# Confidence score (0.0 to 1.0)
confidence = Column(Float, nullable=False, default=0.8, index=True)
# Source tracking: which episodes contributed to this fact
source_episode_ids: Column[list] = Column(
ARRAY(PGUUID(as_uuid=True)), default=list, nullable=False
)
# Learning history
first_learned = Column(DateTime(timezone=True), nullable=False)
last_reinforced = Column(DateTime(timezone=True), nullable=False)
reinforcement_count = Column(Integer, nullable=False, default=1)
# Vector embedding for semantic search
embedding = Column(Vector(1536) if Vector else Text, nullable=True)
# Relationships
project = relationship("Project", foreign_keys=[project_id])
__table_args__ = (
# Unique constraint on triple within project scope
Index(
"ix_facts_unique_triple",
"project_id",
"subject",
"predicate",
"object",
unique=True,
postgresql_where=text("project_id IS NOT NULL"),
),
# Query patterns
Index("ix_facts_subject_predicate", "subject", "predicate"),
Index("ix_facts_project_subject", "project_id", "subject"),
Index("ix_facts_confidence_time", "confidence", "last_reinforced"),
# For finding facts by entity (subject or object)
Index("ix_facts_subject", "subject"),
# Data integrity constraints
CheckConstraint(
"confidence >= 0.0 AND confidence <= 1.0",
name="ck_facts_confidence_range",
),
)
def __repr__(self) -> str:
return (
f"<Fact {self.id} '{self.subject}' - '{self.predicate}' - "
f"'{self.object[:50]}...' conf={self.confidence:.2f}>"
)