diff --git a/openagent_eval/core/pipeline.py b/openagent_eval/core/pipeline.py
index ac0e6c1..ab6e017 100644
--- a/openagent_eval/core/pipeline.py
+++ b/openagent_eval/core/pipeline.py
@@ -12,12 +12,15 @@
 from __future__ import annotations
 
 import inspect
+import logging
 from dataclasses import dataclass, field
 from typing import Any
 
 from openagent_eval.config.models import Config
 from openagent_eval.metrics.base import BaseMetric
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass
 class EvaluationResult:
@@ -193,7 +196,14 @@ def _supports_ground_truth_contexts(self) -> bool:
 
         try:
             sig = inspect.signature(retrieve_method)
-        except (ValueError, TypeError):
+        except (ValueError, TypeError) as e:
+            logger.warning(
+                "Could not inspect signature of %r (%s: %s); "
+                "assuming no ground_truth_contexts support",
+                retrieve_method,
+                type(e).__name__,
+                e,
+            )
             self._retriever_supports_ground_truth_contexts = False
             return False
 
@@ -209,9 +219,21 @@ def _supports_ground_truth_contexts(self) -> bool:
         return False
 
     async def _retrieve(
-        self, question: str, context: str | None, gt_contexts: list[str]
+        self,
+        question: str,
+        context: str | None,
+        gt_contexts: list[str],
     ) -> list[str]:
-        """Retrieve contexts for a question, or fall back to dataset context."""
+        """Retrieve contexts for a question, or fall back to dataset context.
+
+        A retrieval failure is logged (log-only) so a broken retriever is
+        distinguishable from a clean empty retrieval; the fallback behaviour
+        itself is unchanged. The failure is deliberately NOT recorded in
+        ``result.errors``: engine.py derives ``successful_evaluations`` from
+        ``len(results) - len(errors)``, and an entry here has no paired
+        placeholder result, so recording it would mis-count a healthy run
+        as failed.
+        """
         if self._retriever is not None:
             try:
                 if self._supports_ground_truth_contexts():
@@ -221,8 +243,15 @@ async def _retrieve(
                 else:
                     docs = await self._retriever.retrieve(question, k=self._k)
                 return [doc.content for doc in docs]
-            except Exception:
+            except Exception as e:
                 # Retrieval failure -> fall back to any dataset-provided context.
+                logger.warning(
+                    "Retrieval failed for question %r (%s: %s); "
+                    "falling back to dataset-provided context",
+                    question,
+                    type(e).__name__,
+                    e,
+                )
                 if context:
                     return [context]
                 return []
diff --git a/tests/unit/test_core/test_pipeline_retrieval_failure_visibility.py b/tests/unit/test_core/test_pipeline_retrieval_failure_visibility.py
new file mode 100644
index 0000000..3405e81
--- /dev/null
+++ b/tests/unit/test_core/test_pipeline_retrieval_failure_visibility.py
@@ -0,0 +1,95 @@
+"""Regression test for issue #256.
+
+Verifies that a retriever failure inside ``Pipeline._retrieve`` is no longer
+silently swallowed: the dataset-context fallback still applies (behaviour
+unchanged), and the failure is logged naming the exception. The degradation
+is log-only: it must NOT be recorded in the run's ``errors``, because
+engine.py computes ``successful_evaluations`` as ``len(results) -
+len(errors)`` and an entry without a paired placeholder result would
+mis-count a healthy run as failed (breaking cicd gate metrics).
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import pytest
+
+from openagent_eval.config.models import (
+    Config,
+    DatasetConfig,
+    LLMConfig,
+    MetricsConfig,
+    ReportConfig,
+    RetrieverConfig,
+)
+from openagent_eval.core.pipeline import Pipeline
+from openagent_eval.providers.base.retriever import Retriever
+
+if TYPE_CHECKING:
+    from openagent_eval.providers.models import Document
+
+
+class _FailingRetriever(Retriever):
+    """Retriever whose ``retrieve`` always raises, simulating a broken backend."""
+
+    name = "failing"
+    description = "Always raises from retrieve()"
+
+    async def retrieve(
+        self,
+        query: str,
+        k: int = 5,
+        *,
+        ground_truth_contexts: list[str] | None = None,
+    ) -> list[Document]:
+        raise RuntimeError("vector store unreachable")
+
+
+@pytest.mark.asyncio
+async def test_retrieval_failure_is_visible_and_still_falls_back(
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    """A raising retriever degrades to dataset context and is logged only."""
+    from openagent_eval.providers.llm.mock import MockLLMProvider
+
+    config = Config(
+        dataset=DatasetConfig(path="data/questions.json"),
+        llm=LLMConfig(provider="mock", model="mock-model"),
+        retriever=RetrieverConfig(provider="failing"),
+        metrics=MetricsConfig(),
+        report=ReportConfig(),
+        parallel=False,
+    )
+
+    pipeline = Pipeline(
+        config, retriever=_FailingRetriever(), llm=MockLLMProvider()
+    )
+
+    with caplog.at_level(logging.WARNING, logger="openagent_eval.core.pipeline"):
+        result = await pipeline.execute(
+            [
+                {
+                    "question": "What is RAG?",
+                    "ground_truth": "RAG is retrieval augmented generation.",
+                    "context": "dataset-provided context",
+                }
+            ]
+        )
+
+    # (a) Fallback behaviour is unchanged: the dataset context is still used.
+    assert result.results[0].contexts == ["dataset-provided context"]
+
+    # (b) The degradation is log-only: it must NOT appear in result.errors.
+    # engine.py computes successful_evaluations as
+    # len(results) - len(errors), so an errors entry without a paired
+    # placeholder result would mis-count this healthy run as failed and
+    # trip cicd gate metrics. This pins that property.
+    assert result.errors == []
+    assert len(result.results) == 1
+    assert len(result.results) - len(result.errors) == 1
+
+    # (c) The failure is logged, naming the exception.
+    assert "RuntimeError" in caplog.text
+    assert "vector store unreachable" in caplog.text