ca2e28559e
Two files of 1,405 and 814 lines become seven: scope resolution, lifecycle, search, documents, expansion, citations and capabilities. `_config`, `_seed` and the rest move to `helpers.py`, importable by the sandbox tests that share them, and the package points VCR back at `tests/cassettes/multi_db/`.
31 lines
940 B
Python
31 lines
940 B
Python
from pathlib import Path
|
|
|
|
import pytest
|
|
|
|
from haiku.rag.config import get_config
|
|
|
|
|
|
@pytest.fixture(scope="module")
|
|
def vcr_cassette_dir(request):
|
|
"""Cassettes sit with the rest, under `tests/cassettes/multi_db/`."""
|
|
module = request.module.__name__.rsplit(".", 1)[-1]
|
|
return str(Path(__file__).parent.parent / "cassettes" / "multi_db" / module)
|
|
|
|
|
|
@pytest.fixture
|
|
def query_embedding(monkeypatch):
|
|
"""Vector search with no embedder behind it, recording the queries embedded.
|
|
|
|
These tests are about which databases are asked and how often, not about
|
|
retrieval quality, and CI has no embedding endpoint.
|
|
"""
|
|
from haiku.rag.embeddings import EmbedderWrapper
|
|
|
|
embedded: list[str] = []
|
|
|
|
async def embed_query(self, text):
|
|
embedded.append(text)
|
|
return [0.1] * get_config().embeddings.model.vector_dim
|
|
|
|
monkeypatch.setattr(EmbedderWrapper, "embed_query", embed_query)
|
|
return embedded
|