Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

6.1 KB
Raw Blame
"""Embedding dimension mismatch must fail loudly with migration guidance.

The default embedder changed from text-embedding-3-small (1536 dims) to
text-embedding-3-large (3072 dims); stores created before the upgrade must
not silently zero-fill vectors or return empty search results.
"""

from __future__ import annotations

from pathlib import Path
from unittest.mock import MagicMock

import pytest

from crewai.memory.storage.backend import EmbeddingDimensionMismatchError
from crewai.memory.types import MemoryRecord


@pytest.fixture
def lancedb_path(tmp_path: Path) -> Path:
    return tmp_path / "mem"


def _record(dim: int, content: str = "test") -> MemoryRecord:
    return MemoryRecord(content=content, scope="/foo", embedding=[0.1] * dim)


def test_lancedb_save_mismatch_raises(lancedb_path: Path) -> None:
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    storage = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    storage.save([_record(4)])

    with pytest.raises(EmbeddingDimensionMismatchError) as exc_info:
        storage.save([_record(8, "new embedder output")])

    message = str(exc_info.value)
    assert "4-dimensional" in message
    assert "8-dimensional" in message
    assert "crewai reset-memories --memory" in message
    assert "text-embedding-3-small" in message


def test_lancedb_mixed_batch_mismatch_raises(lancedb_path: Path) -> None:
    """A single save() batch with inconsistent dimensions must be rejected."""
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    storage = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    storage.save([_record(4)])

    with pytest.raises(EmbeddingDimensionMismatchError):
        storage.save([_record(4), _record(8, "stray dimension")])


def test_lancedb_mixed_batch_on_fresh_store_raises(lancedb_path: Path) -> None:
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    storage = LanceDBStorage(path=str(lancedb_path))
    with pytest.raises(EmbeddingDimensionMismatchError):
        storage.save([_record(4), _record(8)])


def test_lancedb_search_mismatch_raises(lancedb_path: Path) -> None:
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    storage = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    storage.save([_record(4)])

    with pytest.raises(EmbeddingDimensionMismatchError):
        storage.search([0.1] * 8)


def test_lancedb_update_mismatch_raises(lancedb_path: Path) -> None:
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    storage = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    record = _record(4)
    storage.save([record])

    stale = MemoryRecord(
        id=record.id, content="updated", scope="/foo", embedding=[0.1] * 8
    )
    with pytest.raises(EmbeddingDimensionMismatchError):
        storage.update(stale)


def test_lancedb_reopened_store_detects_mismatch(lancedb_path: Path) -> None:
    """The upgrade scenario: an old store reopened with a new embedder."""
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    old = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    old.save([_record(4)])

    reopened = LanceDBStorage(path=str(lancedb_path))
    with pytest.raises(EmbeddingDimensionMismatchError):
        reopened.save([_record(8)])
    with pytest.raises(EmbeddingDimensionMismatchError):
        reopened.search([0.1] * 8)


def test_memory_reset_all_rebuilds_reopened_store_with_new_dimension(
    lancedb_path: Path,
) -> None:
    from crewai.memory.storage.lancedb_storage import LanceDBStorage
    from crewai.memory.unified_memory import Memory

    old = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    old.save([_record(4)])

    mem = Memory(
        storage=str(lancedb_path),
        llm=MagicMock(),
        embedder=lambda texts: [[0.1] * 8 for _ in texts],
        root_scope="/crew/test",
    )

    mem.reset_all()
    mem.remember(
        "new embedder output",
        scope="/facts",
        categories=["test"],
        importance=0.5,
    )

    assert mem.recall("new embedder output", scope="/facts", depth="shallow")


def test_lancedb_matching_dim_still_works(lancedb_path: Path) -> None:
    from crewai.memory.storage.lancedb_storage import LanceDBStorage

    storage = LanceDBStorage(path=str(lancedb_path), vector_dim=4)
    storage.save([_record(4)])
    storage.save([_record(4, "second")])

    assert len(storage.search([0.1] * 4, limit=5)) == 2


def test_error_is_not_a_runtime_error() -> None:
    """Background-save plumbing treats RuntimeError as executor shutdown and
    silently drops the save — the mismatch must not be classified that way."""
    err = EmbeddingDimensionMismatchError(1536, 3072)
    assert not isinstance(err, RuntimeError)
    assert isinstance(err, ValueError)


def test_background_save_propagates_dimension_mismatch(tmp_path: Path) -> None:
    from unittest.mock import MagicMock

    from crewai.memory.unified_memory import Memory

    mem = Memory(
        storage=str(tmp_path / "db"),
        llm=MagicMock(),
        embedder=lambda texts: [[0.1] * 4 for _ in texts],
    )

    def raise_mismatch(*_args: object, **_kwargs: object) -> None:
        raise EmbeddingDimensionMismatchError(1536, 3072)

    mem._encode_batch = raise_mismatch  # type: ignore[method-assign]

    with pytest.raises(EmbeddingDimensionMismatchError):
        mem._background_encode_batch(["content"], None, None, None, None, None, False, None)


def test_background_save_still_swallows_shutdown_runtime_error(tmp_path: Path) -> None:
    from unittest.mock import MagicMock

    from crewai.memory.unified_memory import Memory

    mem = Memory(
        storage=str(tmp_path / "db"),
        llm=MagicMock(),
        embedder=lambda texts: [[0.1] * 4 for _ in texts],
    )

    def raise_shutdown(*_args: object, **_kwargs: object) -> None:
        raise RuntimeError("cannot schedule new futures after shutdown")

    mem._encode_batch = raise_shutdown  # type: ignore[method-assign]

    assert (
        mem._background_encode_batch(
            ["content"], None, None, None, None, None, False, None
        )
        == []
    )