Fix doctor provider checks: custom endpoints and processing models

haiku_rag_slim/haiku/rag/doctor.py

@@ -83,25 +83,48 @@ def _sample(ids: list[str]) -> list[str]:
     return [*ids[:_SAMPLE_LIMIT], f"... (+{extra} more)"]
 
 
-def _configured_providers(config: AppConfig) -> set[str]:
-    """Providers referenced by the current config across every model role."""
-    providers = {config.embeddings.model.provider}
-    for model in (
-        config.reranking.model,
-        config.qa.model,
-        config.analysis.model,
-    ):
+def _active_models(config: AppConfig) -> list[tuple[str, str, str | None]]:
+    """(provider, name, base_url) for every model role the config activates.
+
+    Picture-description and title models are only included when their feature
+    is enabled (``processing.pictures == "description"`` / ``auto_title``), so
+    doctor checks exactly the providers the next ingest will use.
+    """
+    models = [
+        (
+            config.embeddings.model.provider,
+            config.embeddings.model.name,
+            config.embeddings.model.base_url,
+        )
+    ]
+    for model in (config.reranking.model, config.qa.model, config.analysis.model):
         if model is not None:
-            providers.add(model.provider)
-    return providers
+            models.append((model.provider, model.name, model.base_url))
+
+    proc = config.processing
+    if proc.pictures == "description":
+        pd = proc.conversion_options.picture_description.model
+        models.append((pd.provider, pd.name, pd.base_url))
+    if proc.auto_title:
+        tm = proc.title_model
+        models.append((tm.provider, tm.name, tm.base_url))
+    return models
 
 
 def _check_api_keys(config: AppConfig, environ: dict[str, str]) -> CheckResult:
-    missing: list[str] = []
-    for provider in sorted(_configured_providers(config)):
-        env_var = _PROVIDER_ENV_VARS.get(provider)
-        if env_var and not environ.get(env_var):
-            missing.append(f"{provider} ({env_var})")
+    # A custom base_url points at a self-hosted OpenAI-compatible endpoint that
+    # uses a placeholder key, so the SaaS key is only required when a provider
+    # is used without one. Reachability of custom endpoints is the probe's job.
+    need_key = {
+        provider
+        for provider, _name, base_url in _active_models(config)
+        if not base_url and provider in _PROVIDER_ENV_VARS
+    }
+    missing = [
+        f"{provider} ({_PROVIDER_ENV_VARS[provider]})"
+        for provider in sorted(need_key)
+        if not environ.get(_PROVIDER_ENV_VARS[provider])
+    ]
     if missing:
         return CheckResult(
             name="api_keys",
@@ -607,14 +630,8 @@ def _provider_targets(
                 {"kind": "docling-serve", "display": base, "models": set()},
             )
 
-    add_model(
-        config.embeddings.model.provider,
-        config.embeddings.model.name,
-        config.embeddings.model.base_url,
-    )
-    for model in (config.reranking.model, config.qa.model, config.analysis.model):
-        if model is not None:
-            add_model(model.provider, model.name, model.base_url)
+    for provider, name, base_url in _active_models(config):
+        add_model(provider, name, base_url)
 
     return targets, local
 

tests/test_doctor.py

@@ -9,9 +9,12 @@ from typer.testing import CliRunner
 from haiku.rag.cli import _cli as cli
 from haiku.rag.config.models import (
     AppConfig,
+    ConversionOptions,
     DoclingServeConfig,
     EmbeddingModelConfig,
     EmbeddingsConfig,
+    ModelConfig,
+    PictureDescriptionConfig,
     ProcessingConfig,
     ProvidersConfig,
 )
@@ -19,6 +22,8 @@ from haiku.rag.doctor import (
     CheckResult,
     DoctorReport,
     Severity,
+    _active_models,
+    _check_api_keys,
     _check_embedding_drift,
     _check_vector_index,
     _model_present,
@@ -628,6 +633,67 @@ def test_cli_doctor_exits_1_on_failure(monkeypatch):
     assert result.exit_code == 1
 
 
+# --- Active models / API keys ---
+
+
+def test_api_key_not_required_for_custom_openai_base_url():
+    config = AppConfig(
+        embeddings=EmbeddingsConfig(
+            model=EmbeddingModelConfig(
+                provider="openai",
+                name="x",
+                vector_dim=4,
+                base_url="http://localhost:1234/v1",
+            )
+        )
+    )
+    assert _check_api_keys(config, {}).severity is Severity.OK
+
+
+def test_api_key_required_for_openai_without_base_url():
+    config = AppConfig(
+        embeddings=EmbeddingsConfig(
+            model=EmbeddingModelConfig(provider="openai", name="x", vector_dim=4)
+        )
+    )
+    result = _check_api_keys(config, {})
+    assert result.severity is Severity.FAIL
+    assert any("OPENAI_API_KEY" in d for d in result.details)
+
+
+def test_active_models_includes_picture_description_when_enabled():
+    config = AppConfig(processing=ProcessingConfig(pictures="description"))
+    names = [name for _p, name, _b in _active_models(config)]
+    assert "ministral-3" in names
+
+
+def test_active_models_excludes_picture_description_by_default():
+    names = [name for _p, name, _b in _active_models(AppConfig())]
+    assert "ministral-3" not in names
+
+
+def test_active_models_includes_title_model_when_auto_title():
+    base = _active_models(AppConfig())
+    with_title = _active_models(AppConfig(processing=ProcessingConfig(auto_title=True)))
+    assert len(with_title) == len(base) + 1
+
+
+def test_picture_description_model_checked_for_api_key():
+    config = AppConfig(
+        processing=ProcessingConfig(
+            pictures="description",
+            conversion_options=ConversionOptions(
+                picture_description=PictureDescriptionConfig(
+                    model=ModelConfig(provider="openai", name="gpt-4o")
+                )
+            ),
+        )
+    )
+    result = _check_api_keys(config, {})
+    assert result.severity is Severity.FAIL
+    assert any("OPENAI_API_KEY" in d for d in result.details)
+
+
 # --- Provider connectivity ---