docs: discussion recipe models relation

docs/proposals/recipe-models-relation.md

@@ -0,0 +1,70 @@
+## Recipes
+
+Currently recipes cannot uses imported models, as they hardcode all models they are compatible with
+
+Example with the chatbot recipe
+````json
+{
+  "id": "chatbot",
+  ...
+  "categories": [
+    "natural-language-processing"
+  ],
+  "models": [
+    "hf.instructlab.granite-7b-lab-GGUF",
+    "hf.instructlab.merlinite-7b-lab-GGUF",
+    ...
+  ]
+}
+````
+
+We want to find a solution to allow user to use their imported models.
+
+## Models <=> Recipes relation
+
+We discussed earlier, that having a direct relation between models and recipe is 
+not the best, as it forces us to hardcode one of them.
+
+The solution would be to introduce a new _element_, which would act as an intermediate.
+
+### E.g.
+
+Models and recipes would define the `backends? | providers? | categories?` they can run on, and the recipe would also define the same. 
+
+Therefore, we would be able to add dynamically models and recipes without having to hardcode anything.
+
+### Problem
+
+**How to name the element linking those**, with https://github.com/containers/podman-desktop-extension-ai-lab/pull/1161 I introduced `InferenceProvider`.
+
+However, currently they are only defined by a name, and this is not really ideal.
+
+Let's say in the future I have several InferenceProviders
+
+- llama-cpp
+- llama-cpp-cuda
+- llama-cpp-amd
+- ollama
+- ollama-cuda
+- ollama-romc
+- whisper-cpp
+- whisper-cuda
+- whisper-amd
+
+how do I define them ? how recipes and models can indicate that they support one or another ?
+
+For the `llama-cpp*` and `ollama*` they technically are the same logic (llamacpp) they can run `large-language-models`.
+
+Therefore we might use the existing categories ?
+
+Should we implement each InferenceProviders individually ?
+
+Should we add a property to providers to group them ? (tags, labels, implementation)
+
+### Personal
+
+I am having difficulty choosing the right wording to define them.
+
+An easy solution might be to add a `type: string` property to InferenceProvider.
+
+And uses `backend: string` for models and recipe, and llama-cpp would be the type for all llama-cpp* and ollama*