Add llmman as a self-hosted host for chat and embedding providers (#947)

Co-authored-by: Thorsten Sommer <SommerEngineering@users.noreply.github.com>

app/Tests/Models/Corpus/ModelKindCorpus.cs

@@ -66,6 +66,14 @@ public static class ModelKindCorpus
         new(SELF_HOSTED, "LaBSE", EMBEDDING),
         new(SELF_HOSTED, "instructor-xl", EMBEDDING),
         new(SELF_HOSTED, "gtr-t5-large", EMBEDDING),
+
+        //
+        // llmman lists every model by the full reference it pulled it from: a registry, a namespace,
+        // the repository, and a tag. Whatever is in front of the model has to come off before the
+        // word which says what it is for can be found:
+        //
+        new(SELF_HOSTED, "docker.io/ai/embeddinggemma:latest", EMBEDDING),
+        new(SELF_HOSTED, "docker.io/ai/nomic-embed-text-v1.5:latest", EMBEDDING),
     ];
 
     /// <summary>
@@ -144,6 +152,9 @@ public static class ModelKindCorpus
         new(ALIBABA_CLOUD, "qwen3-asr-flash", TRANSCRIPTION),
         new(ALIBABA_CLOUD, "qwen3-asr-1.7b", TRANSCRIPTION),
         new(ALIBABA_CLOUD, "qwen-audio-3.0-asr-flash-streaming", TRANSCRIPTION),
+
+        // Voxtral as llmman lists it, from the GGUF build on Hugging Face which llama.cpp serves:
+        new(SELF_HOSTED, "hf.co/ggml-org/Voxtral-Mini-3B-2507-GGUF:latest", TRANSCRIPTION),
     ];
 
     /// <summary>
@@ -395,6 +406,14 @@ public static class ModelKindCorpus
         new(GROQ, "groq/compound", CHAT),
         new(GROQ, "groq/compound-mini", CHAT),
         new(GOOGLE, "gemini-robotics-er-1.5-preview", CHAT),
+
+        //
+        // Chat models as llmman lists them, with the registry, the namespace, and the tag still in
+        // front of and behind the model. None of these parts may turn one into something else:
+        //
+        new(SELF_HOSTED, "hf.co/unsloth/Qwen3.5-0.8B-GGUF:latest", CHAT),
+        new(SELF_HOSTED, "docker.io/ai/qwen3.5:0.8b", CHAT),
+        new(SELF_HOSTED, "hf.co/HuggingFaceTB/SmolLM2-135M-Instruct:latest", CHAT),
     ];
 
     /// <summary>

app/Tests/Models/Hosting/ModelHostTests.cs

@@ -172,6 +172,25 @@ public sealed class ModelHostTests
         Assert.That(unwrapped.Original, Is.EqualTo("qwen3.8:27b-mlx"));
     }
 
+    [TestCase("hf.co/HuggingFaceTB/SmolLM2-135M-Instruct:latest", "SmolLM2-135M-Instruct:latest", ModelVendor.HUGGING_FACE, TestName = "Behind the registry, the organization from the hub names the vendor")]
+    [TestCase("hf.co/unsloth/Qwen3.5-0.8B-GGUF:latest", "Qwen3.5-0.8B-GGUF:latest", null, TestName = "An organization which only repackaged the weights names nobody")]
+    [TestCase("docker.io/ai/qwen3.5:0.8b", "qwen3.5:0.8b", null, TestName = "The registry and its namespace come off and the variant stays")]
+    public void AnLlmmanReferenceLosesItsRegistryAndKeepsItsTag(string modelId, string expectedModel, ModelVendor? expectedVendor)
+    {
+        //
+        // llmman lists the reference it pulled a model from, with the registry in front and the tag
+        // behind. The registry is a path segment like any organization and comes off the same way.
+        // The tag is what Ollama writes after its colon, too, and stays for the same reason.
+        //
+        var unwrapped = Unwrap(LLMProviders.SELF_HOSTED, modelId, out var vendor);
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(unwrapped.Original, Is.EqualTo(expectedModel));
+            Assert.That(vendor, Is.EqualTo(expectedVendor));
+        });
+    }
+
     [Test]
     public void AResellerLeavesTheNameAloneAndOnlyTakesTheApiAway()
     {