OllamaModelBackend.is_model_available() is case-sensitive, causing false negatives against locally installed models

[lamoghaighoroje]

Description

OllamaModelBackend.is_model_available() (mellea/backends/ollama.py) compares model names case-sensitively via Python's str.startswith(), but Ollama itself treats model tags case-insensitively everywhere else (CLI ollama show, the HTTP /api/show endpoint, ollama pull, ollama cp). This causes is_model_available() to return False for a model that is genuinely installed and resolvable, whenever the locally-registered tag's casing differs from the casing used in the check — which can easily diverge because Ollama silently canonicalizes/normalizes a model's tag casing on pull, independent of what the user types.

Steps to Reproduce

1. Have Ollama register a model under a tag with mixed case, e.g. granite4.1:3B (confirmed via ollama list, ollama pull granite4.1:3b, and ollama cp granite4.1:3B granite4.1:3b all normalizing to/keeping the same 3B entry — Ollama does not create a distinct lowercase entry no matter how the tag is typed).
2. Call is_model_available("granite4.1:3b") (lowercase, matching how a config elsewhere in your own stack might store it) against that same install.

from mellea.backends.ollama import OllamaModelBackend
from ollama import Client

b = OllamaModelBackend.__new__(OllamaModelBackend)
b._client = Client()
print(b.is_model_available("granite4.1:3b"))   # False
print(b.is_model_available("granite4.1:3B"))   # True

Meanwhile, Ollama's own tooling resolves the lowercase form fine:

$ ollama show granite4.1:3b   # exits 0, prints model details

And the HTTP API agrees:

$ curl -s http://localhost:11434/api/show -d '{"model": "granite4.1:3b"}'
# returns a normal 200 response with model details

So ollama show / /api/show resolve case-insensitively server-side, but is_model_available()'s client-side comparison against the list returned by /api/tags does not.

Root Cause

mellea/backends/ollama.py, is_model_available():

def is_model_available(self, model_name):
    ...
    try:
        models = self._client.list()
        for model in models["models"]:
            if model.model.startswith(model_name):
                return True
        return False
    except Exception as e:
        print(f"An error occurred: {e}")
        return False

models["models"][i].model is whatever casing Ollama's /api/tags returns for the locally registered tag (its canonical casing, which the caller does not control). Comparing that against a caller-supplied model_name with str.startswith() is case-sensitive, so any casing mismatch between the two — even though both refer to the exact same model as far as Ollama itself is concerned — produces a False.

This also affects _pull_ollama_model(), since it short-circuits on is_model_available(self._model_id) before attempting a pull, and can surface downstream as "Model '<name>' not found. Please download it first." even when the model is already installed under a differently-cased tag.

Confirmed still present in v0.8.0 and on main

Checked directly against the v0.8.0 tag and the current main branch — the method is unchanged (only a self._raise_if_closed() guard was added ahead of the comparison logic; the case-sensitive startswith() check itself is identical).

Proposed Fix

Case-fold both sides before comparing, matching how Ollama's own CLI/HTTP API already resolve tags:

for model in models["models"]:
    if model.model.lower().startswith(model_name.lower()):
        return True

Environment

• mellea version: confirmed present in 0.7.0, 0.8.0, and current main
• Ollama: local install, model registered as granite4.1:3B
• macOS (M5 Max)