feat: implement GEWU service and local agent workflows

tests/adapters/test_embeddings.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+import hashlib
+from pathlib import Path
+from typing import Any
+
+import pytest
+
+from gewu.adapters.embeddings import (
+    EmbeddingProviderError,
+    LocalSentenceTransformer,
+    model_artifact_sha256,
+)
+
+
+class FakeEncoder:
+    def __init__(self, rows: list[list[float]]) -> None:
+        self.rows = rows
+        self.seen: list[str] = []
+        self.calls = 0
+
+    def encode(
+        self,
+        sentences: list[str],
+        *,
+        normalize_embeddings: bool,
+        convert_to_numpy: bool,
+        show_progress_bar: bool,
+    ) -> Any:
+        assert normalize_embeddings and convert_to_numpy and not show_progress_bar
+        self.seen.extend(sentences)
+        result = self.rows[self.calls : self.calls + len(sentences)]
+        self.calls += len(sentences)
+        return result
+
+
+def _model_path(tmp_path: Path) -> Path:
+    path = tmp_path / "model"
+    path.mkdir()
+    (path / "config.json").write_text('{"dim":2}', encoding="utf-8")
+    return path
+
+
+def test_local_sentence_transformer_applies_e5_prefixes_and_metadata(tmp_path: Path) -> None:
+    path = _model_path(tmp_path)
+    encoder = FakeEncoder([[0.6, 0.8], [1.0, 0.0]])
+    artifact = model_artifact_sha256(path)
+    provider = LocalSentenceTransformer(
+        path,
+        model_id="intfloat/multilingual-e5-small@pinned",
+        expected_sha256=artifact,
+        dimensions=2,
+        encoder=encoder,
+    )
+
+    assert provider.embed_queries(("question",)) == ((0.6, 0.8),)
+    assert provider.embed_passages(("evidence",)) == ((1.0, 0.0),)
+    assert encoder.seen == ["query: question", "passage: evidence"]
+    assert provider.model_id == "intfloat/multilingual-e5-small@pinned"
+    assert provider.artifact_sha256 == artifact
+    assert len(provider.artifact_sha256) == 64
+
+
+def test_local_sentence_transformer_rejects_changed_artifact(tmp_path: Path) -> None:
+    path = _model_path(tmp_path)
+    (path / "config.json").write_text("changed", encoding="utf-8")
+
+    with pytest.raises(EmbeddingProviderError, match="digest"):
+        LocalSentenceTransformer(
+            path,
+            model_id="model",
+            expected_sha256=hashlib.sha256(b"wrong").hexdigest(),
+            dimensions=2,
+            encoder=FakeEncoder([[1.0, 0.0]]),
+        )
+
+
+@pytest.mark.parametrize("rows", [[[2.0, 0.0]], [[float("nan"), 0.0]], [[1.0]]])
+def test_local_sentence_transformer_rejects_invalid_vectors(
+    tmp_path: Path, rows: list[list[float]]
+) -> None:
+    path = _model_path(tmp_path)
+    provider = LocalSentenceTransformer(
+        path,
+        model_id="model",
+        expected_sha256=model_artifact_sha256(path),
+        dimensions=2,
+        encoder=FakeEncoder(rows),
+    )
+
+    with pytest.raises(EmbeddingProviderError):
+        provider.embed_passages(("text",))