diff --git a/CHANGELOG.md b/CHANGELOG.md
index 816cf1f8..2f5a0752 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Changed
+
+- Chat model provider `gemini` renamed to `google`, matching pydantic-ai. Update `provider: gemini` to `provider: google`.
+
 ### Added
 
 - `lancedb.databases` configures a named set of local or remote databases.
@@ -33,6 +37,7 @@
 - Inspector search results mark truncated previews with an ellipsis.
 - Document titles, URIs, headings and database names render as text, not Rich
   markup, in `search` output, chat citations and the chat document filter.
+- An unrecognized chat model provider raises `Unknown model provider '<name>'` instead of reaching pydantic-ai as a `provider:name` string, and outranks the `api_key` check, so an unusable provider is no longer reported as a missing vendor environment variable.
 
 ## [0.78.0] - 2026-08-24
 
diff --git a/docs/configuration/providers.md b/docs/configuration/providers.md
index 90d76f26..5111dd33 100644
--- a/docs/configuration/providers.md
+++ b/docs/configuration/providers.md
@@ -127,7 +127,7 @@ qa:
 
 Same mechanism, opposite direction. Without `extra_body` the Gemma-4 chat template defaults to non-thinking and dumps a verbose answer straight into `content`. With it on, vLLM (started with `--reasoning-parser`) populates the parsed `reasoning` field and leaves `content` as the concise final answer.
 
-**Provider support:** honored by openai, ollama, anthropic, and groq via pydantic-ai's `ModelSettings.extra_body`. Silently ignored by gemini and bedrock.
+**Provider support:** honored by openai, ollama, anthropic, and groq via pydantic-ai's `ModelSettings.extra_body`. Silently ignored by google and bedrock.
 
 ## Embedding Providers
 
@@ -393,7 +393,7 @@ Any provider supported by Pydantic AI can be used. Examples:
 # Google Gemini
 qa:
   model:
-    provider: gemini
+    provider: google
     name: gemini-1.5-flash
 
 # Groq
diff --git a/haiku_rag_slim/haiku/rag/config/models.py b/haiku_rag_slim/haiku/rag/config/models.py
index e019ecd5..a8a938c5 100644
--- a/haiku_rag_slim/haiku/rag/config/models.py
+++ b/haiku_rag_slim/haiku/rag/config/models.py
@@ -36,7 +36,7 @@ class ModelConfig(ConfigModel):
             `ModelSettings.extra_body`. Provider-side escape hatch for
             keys haiku.rag doesn't model explicitly (e.g. vLLM's
             `chat_template_kwargs.enable_thinking: false` for Qwen3).
-            Honored by openai/ollama/anthropic/groq; ignored by gemini/bedrock.
+            Honored by openai/ollama/anthropic/groq; ignored by google/bedrock.
     """
 
     provider: str = "ollama"
diff --git a/haiku_rag_slim/haiku/rag/utils.py b/haiku_rag_slim/haiku/rag/utils.py
index 8e115f02..125da1fc 100644
--- a/haiku_rag_slim/haiku/rag/utils.py
+++ b/haiku_rag_slim/haiku/rag/utils.py
@@ -62,6 +62,32 @@ def check_api_key_supported(
         )
 
 
+def _check_provider_known(provider: str) -> None:
+    """Reject a chat provider pydantic-ai cannot resolve.
+
+    Providers we do not branch on reach pydantic-ai as a `provider:name` string,
+    so without this an unusable name fails deep inside pydantic-ai with nothing
+    naming the config key it came from. Asking pydantic-ai's own resolver rather
+    than keeping a list here leaves a newly added provider working with no
+    release of ours.
+    """
+    from pydantic_ai.providers import infer_provider_class
+
+    try:
+        infer_provider_class(provider)
+    except ImportError:
+        # pydantic-ai knows the name, its SDK is just not installed here. That
+        # failure names the extra to install, so leave it to be raised in place.
+        return
+    except ValueError:
+        raise ValueError(
+            f"Unknown model provider '{provider}'. See "
+            "https://ai.pydantic.dev/models/ for the providers pydantic-ai "
+            "supports. An OpenAI-compatible server (vLLM, sglang, LM Studio) "
+            "uses provider 'openai' with base_url."
+        ) from None
+
+
 def cosine_similarity(vec1: list[float], vec2: list[float]) -> float:
     """Compute cosine similarity between two vectors."""
     dot_product = sum(a * b for a, b in zip(vec1, vec2))
@@ -169,6 +195,7 @@ def get_model(
 
     provider = model_config.provider
     model = model_config.name
+    _check_provider_known(provider)
     check_api_key_supported(model_config, {"openai", "ollama"})
 
     if provider == "ollama":
@@ -261,7 +288,7 @@ def get_model(
 
         return AnthropicModel(model_name=model, settings=anthropic_settings)
 
-    elif provider == "gemini":
+    elif provider == "google":
         from pydantic_ai.models.google import GoogleModel
 
         return GoogleModel(
diff --git a/tests/test_utils.py b/tests/test_utils.py
index 0a48950f..92224dd3 100644
--- a/tests/test_utils.py
+++ b/tests/test_utils.py
@@ -390,23 +390,23 @@ def test_get_model_anthropic_thinking_off_disables_adaptive_models():
 
 
 @pytest.mark.skipif(not HAS_GOOGLE, reason="Google not installed")
-def test_get_model_gemini():
-    """Test get_model returns GoogleModel for Gemini."""
+def test_get_model_google():
+    """Test get_model returns GoogleModel for Google."""
     from pydantic_ai.models.google import GoogleModel
 
-    model_config = ModelConfig(provider="gemini", name="gemini-2.0-flash-exp")
+    model_config = ModelConfig(provider="google", name="gemini-2.0-flash-exp")
     result = get_model(model_config)
     assert isinstance(result, GoogleModel)
 
 
 @pytest.mark.skipif(not HAS_GOOGLE, reason="Google not installed")
 @pytest.mark.parametrize("enable_thinking", [True, False])
-def test_get_model_gemini_with_thinking(enable_thinking):
-    """Test get_model configures thinking for Gemini."""
+def test_get_model_google_with_thinking(enable_thinking):
+    """Test get_model configures thinking for Google."""
     from pydantic_ai.models.google import GoogleModel
 
     model_config = ModelConfig(
-        provider="gemini",
+        provider="google",
         name="gemini-2.0-flash-thinking-exp",
         enable_thinking=enable_thinking,
     )
@@ -536,14 +536,45 @@ def test_get_model_bedrock_rejects_mantle_only_model():
         get_model(model_config)
 
 
-def test_get_model_unknown_provider():
-    """Test get_model returns string format for unknown providers."""
+def test_get_model_passthrough_for_unbranched_provider():
+    """A provider pydantic-ai knows but we do not branch on passes through as a
+    string, so a new pydantic-ai provider needs no haiku.rag release."""
     model_config = ModelConfig(provider="mistral", name="mistral-large-latest")
     result = get_model(model_config)
     assert isinstance(result, str)
     assert result == "mistral:mistral-large-latest"
 
 
+def test_get_model_accepts_provider_whose_sdk_is_missing(monkeypatch):
+    """A missing vendor SDK is not an unknown provider: that ImportError names
+    the extra to install, so it must reach the caller unchanged.
+
+    Uses a provider whose SDK *is* installed, so the patch is what produces the
+    ImportError rather than the environment.
+    """
+    import pydantic_ai.providers
+
+    def _missing_sdk(provider: str):
+        raise ImportError("Please install the `cohere` package")
+
+    monkeypatch.setattr(pydantic_ai.providers, "infer_provider_class", _missing_sdk)
+    result = get_model(ModelConfig(provider="cohere", name="command-r"))
+
+    assert result == "cohere:command-r"
+
+
+@pytest.mark.parametrize("provider", ["nonsense", "vllm", "gemini"])
+def test_get_model_rejects_unknown_provider(provider):
+    """An unknown provider is named here rather than passed through to fail
+    inside pydantic-ai, where nothing identifies the config it came from.
+
+    `vllm` and `gemini` get no special case: both were haiku.rag's own
+    vocabulary, and both fail the same way as a typo.
+    """
+    with pytest.raises(ValueError, match=provider):
+        get_model(ModelConfig(provider=provider, name="whatever"))
+
+
 def test_get_package_versions():
     """Test get_package_versions returns expected keys."""
     from haiku.rag.utils import get_package_versions