fix(embedding): stop quota failures and transient item errors from multiplying

Review follow-up on the two P2 findings in #579:

* quota exhaustion no longer falls into the per-item split. The store now gives
  up explicitly when the provider reports insufficient_quota, including after
  the opt-in quota retries are spent, because no single item is to blame and
  splitting only multiplied the rejected requests (3 items -> 4 requests).
* the per-item fallback goes back through `_call_with_retry`, so a 429,
  connection error or timeout on an isolated item keeps the existing backoff
  and retry policy instead of silently dropping that item's vector. A one-item
  request still cannot split again.

Regression tests cover both findings: a quota failure stays a single request
with the default policy and with exhausted quota retries, and a transient
failure on one item of a split batch still ends with every healthy sibling
embedded.

python -m pytest tests/unit/test_local_embedding_store.py tests/unit/test_file_store_consistency.py -q
-> 83 passed, 41 skipped

reme/components/embedding_store/local_embedding_store.py

@@ -213,6 +213,13 @@ class LocalEmbeddingStore(BaseEmbeddingStore):
                     )
                     await asyncio.sleep(self.quota_retry_delay)
                     continue
+                if self._is_insufficient_quota(error):
+                    # Quota is a service-level failure: no single item is to blame, so splitting the
+                    # batch cannot isolate anything and would only multiply the rejected requests.
+                    # This also covers the case where the quota retries above are exhausted.
+                    self.logger.exception("Embedding request failed: quota exhausted")
+                    self.is_healthy = False
+                    return None
                 self.logger.exception("Embedding request failed")
                 self.is_healthy = False
                 if len(texts) > 1:
@@ -227,19 +234,18 @@ class LocalEmbeddingStore(BaseEmbeddingStore):
         Provider-side request errors, such as a per-item context-length limit, reject the
         whole batch and would otherwise cost every vector in it. Isolating each item bounds
         the damage to the item that actually failed and keeps the remaining vectors usable.
+        Each single-item request goes back through ``_call_with_retry`` so a transient
+        failure (429, connection error, timeout) still gets the existing backoff and retry
+        policy instead of losing the vector; a one-item request never splits again.
         """
         results: list[list[float] | None] = []
         for text in texts:
-            try:
+            single = await self._call_with_retry([text], **kwargs)
-                result = await self.as_embedding([text], **kwargs)
+            if single and len(single) == 1:
-            except Exception as error:
+                results.append(single[0])
-                self.logger.error(
-                    f"Embedding request failed for a single item (chars={len(text)}): "
-                    f"{type(error).__name__}: {error}",
-                )
-                results.append(None)
                 continue
-            results.append(result[0] if result and len(result) == 1 else None)
+            self.logger.error(f"Embedding request failed for a single item (chars={len(text)})")
+            results.append(None)
         return results
 
     @staticmethod

tests/unit/test_local_embedding_store.py

@@ -158,6 +158,48 @@ class PerItemLimitAsEmbedding:
         return [[1.0, 0.0] for _ in texts]
 
 
+class AlwaysInsufficientQuotaAsEmbedding:
+    """Always fail with quota exhaustion, recording every requested batch size."""
+
+    dimensions = 2
+
+    def __init__(self):
+        self.batch_sizes: list[int] = []
+
+    async def __call__(self, texts: list[str], **_kwargs):
+        self.batch_sizes.append(len(texts))
+        raise InsufficientQuotaError("quota exhausted")
+
+
+class PerItemLimitThenTransientErrorAsEmbedding:
+    """Reject an over-limit batch, then fail the first single-item request once.
+
+    Mirrors the review case: the batch is rejected because one item exceeds the per-item
+    window, and a sibling item then hits a transient provider failure while retried alone.
+    """
+
+    dimensions = 2
+    vector_space_id = "fakespace000"
+    max_chars = 32
+
+    def __init__(self, transient: Exception):
+        self.transient = transient
+        self.batch_sizes: list[int] = []
+        self.transient_failures = 0
+
+    def initialize_model(self):
+        """Mirror the real component's idempotent initialization hook."""
+
+    async def __call__(self, texts: list[str], **_kwargs):
+        self.batch_sizes.append(len(texts))
+        if any(len(text) > self.max_chars for text in texts):
+            raise ContextLengthError("the input length exceeds the context length")
+        if len(texts) == 1 and self.transient_failures == 0:
+            self.transient_failures += 1
+            raise self.transient
+        return [[1.0, 0.0] for _ in texts]
+
+
 class BadNodeEmbeddingStore(BaseEmbeddingStore):
     """Embedding store that returns wrong-dimensional vectors."""
 
@@ -828,3 +870,84 @@ def test_rate_limited_batch_is_not_split_into_per_item_requests(monkeypatch):
         assert sleeps == [1.0]
 
     run(go())
+
+
+def test_insufficient_quota_batch_is_not_split_into_per_item_requests(monkeypatch):
+    """Quota exhaustion is a service-level failure; splitting it would multiply the load."""
+
+    async def go():
+        sleeps = []
+
+        async def fake_sleep(delay):
+            sleeps.append(delay)
+
+        store = LocalEmbeddingStore(name="t_local_embedding_quota_batch", max_retries=2)
+        embedding = AlwaysInsufficientQuotaAsEmbedding()
+        store.as_embedding = embedding
+        monkeypatch.setattr(asyncio, "sleep", fake_sleep)
+
+        assert await store._call_with_retry(["a", "b", "c"]) is None
+        assert embedding.batch_sizes == [3]
+        assert store.is_healthy is False
+        assert not sleeps
+
+    run(go())
+
+
+def test_exhausted_quota_retries_still_do_not_split_the_batch(monkeypatch):
+    """Once the opt-in quota retries run out, the store gives up instead of splitting."""
+
+    async def go():
+        sleeps = []
+
+        async def fake_sleep(delay):
+            sleeps.append(delay)
+
+        store = LocalEmbeddingStore(
+            name="t_local_embedding_quota_batch_exhausted",
+            max_retries=3,
+            quota_retry_delay=5.0,
+        )
+        embedding = AlwaysInsufficientQuotaAsEmbedding()
+        store.as_embedding = embedding
+        monkeypatch.setattr(asyncio, "sleep", fake_sleep)
+
+        assert await store._call_with_retry(["a", "b"]) is None
+        assert embedding.batch_sizes == [2, 2, 2]
+        assert sleeps == [5.0, 5.0]
+
+    run(go())
+
+
+def test_per_item_fallback_keeps_a_transiently_failed_vector(monkeypatch):
+    """A 429 / connection error / timeout on one isolated item must not lose its vector."""
+
+    async def go():
+        transients = (
+            RateLimitError("Requests are too frequent"),
+            ConnectionError("connection reset"),
+            TimeoutError("request timed out"),
+        )
+        for transient in transients:
+            sleeps = []
+
+            async def fake_sleep(delay):
+                sleeps.append(delay)
+
+            provider = PerItemLimitThenTransientErrorAsEmbedding(transient)
+            store = LocalEmbeddingStore(
+                name="t_local_embedding_fallback_transient",
+                max_retries=3,
+                enable_cache=False,
+            )
+            store.as_embedding = provider
+            monkeypatch.setattr(asyncio, "sleep", fake_sleep)
+
+            results = await store._call_with_retry(["ok", "x" * 64, "also ok"])
+
+            assert results == [[1.0, 0.0], None, [1.0, 0.0]]
+            assert provider.batch_sizes == [3, 1, 1, 1, 1]
+            assert provider.transient_failures == 1
+            assert sleeps == [1.0]
+
+    run(go())