feat: ollama embedder mit exponential backoff retry

tests/test_embedder.py

@@ -0,0 +1,52 @@
+import asyncio
+from unittest.mock import AsyncMock, MagicMock, patch
+import pytest
+
+from app.ingest.embedder import embed_texts, EmbeddingError
+
+
+@pytest.mark.asyncio
+async def test_embed_texts_returns_vectors():
+    fake_client = MagicMock()
+    fake_client.embeddings = AsyncMock(side_effect=[
+        {"embedding": [0.1, 0.2, 0.3]},
+        {"embedding": [0.4, 0.5, 0.6]},
+    ])
+
+    with patch("app.ingest.embedder._client", return_value=fake_client):
+        vectors = await embed_texts(["hello", "world"], model="qwen3-embedding:0.6b")
+
+    assert vectors == [[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]]
+    assert fake_client.embeddings.call_count == 2
+
+
+@pytest.mark.asyncio
+async def test_embed_texts_retries_on_failure(monkeypatch):
+    monkeypatch.setattr("app.ingest.embedder._BACKOFF_SECONDS", [0, 0, 0])
+
+    fake_client = MagicMock()
+    fake_client.embeddings = AsyncMock(side_effect=[
+        Exception("connection refused"),
+        Exception("timeout"),
+        {"embedding": [1.0, 2.0]},
+    ])
+
+    with patch("app.ingest.embedder._client", return_value=fake_client):
+        vectors = await embed_texts(["hi"], model="m")
+
+    assert vectors == [[1.0, 2.0]]
+    assert fake_client.embeddings.call_count == 3
+
+
+@pytest.mark.asyncio
+async def test_embed_texts_raises_after_max_retries(monkeypatch):
+    monkeypatch.setattr("app.ingest.embedder._BACKOFF_SECONDS", [0, 0, 0])
+
+    fake_client = MagicMock()
+    fake_client.embeddings = AsyncMock(side_effect=Exception("nope"))
+
+    with patch("app.ingest.embedder._client", return_value=fake_client):
+        with pytest.raises(EmbeddingError):
+            await embed_texts(["hi"], model="m")
+
+    assert fake_client.embeddings.call_count == 4  # initial + 3 retries