[Feature]: Google Gemini as AI provider

[RobertSmits]

Problem Statement

At the moment only OpenAI and LocalAI are officialy supported.
I would like to use Google Gemini as the AI provider.
This can be done by using their OpenAI compatible API: https://generativelanguage.googleapis.com/v1beta/openai/
This however causes issues as Gemini doesn't support logprobs used in the tag request call.

Proposed Solution

If the use of logprobs in the AI calls can be made optional based on AI provider, this issue would be resolved.

Alternatives Considered

No response

Feature Area

AI/Outfit Recommendations

How important is this feature to you?

Nice to have

Additional Context

logprobs only gets used to get tags for clothing items.
As a test I set request_logprobs to false for this call and AI functions now seem to work correctly with Gemini

Contribution

• I would be willing to help implement this feature

[virusz4274]

I was also curious with using Google models on this. When scouring through the code i found that if we add a retry logic without the logprobs we can have gemini working.

AIService._call_with_fallback() currently requests logprobs during vision tagging. Some providers reject logprobs/top_logprobs, so the retry loop repeats the same failing request. A better default is to retry once without logprobs when the provider returns a 400/422 unsupported-parameter error.

Suggested helper:

def _is_unsupported_logprobs_error(error: httpx.HTTPStatusError) -> bool:
    if error.response.status_code not in (400, 422):
        return False

    body = error.response.text.lower()
    return (
        "logprobs" in body or "top_logprobs" in body
    ) and (
        "unsupported" in body
        or "unrecognized" in body
        or "unknown"
        or "invalid" in body
    )

Then in _call_with_fallback(), track whether logprobs are still enabled:

logprobs_supported = request_logprobs

for attempt in range(self.settings.ai_max_retries):
    try:
        request_body = {
            "model": model,
            "messages": messages,
            "stream": False,
            "max_tokens": self.settings.ai_max_tokens,
        }

        if logprobs_supported:
            request_body["logprobs"] = True
            request_body["top_logprobs"] = 3

        response = await client.post(
            f"{endpoint.url}/chat/completions",
            headers=self._get_headers(),
            json=request_body,
        )
        response.raise_for_status()

        data = response.json()
        choice = data["choices"][0]
        content = choice["message"]["content"]

        logprobs_content = None
        if logprobs_supported and choice.get("logprobs"):
            logprobs_content = choice["logprobs"].get("content")

        return content, None, logprobs_content

    except httpx.HTTPStatusError as e:
        if logprobs_supported and _is_unsupported_logprobs_error(e):
            logger.info(
                f"{endpoint.name} does not support logprobs; retrying {task_name} without logprobs"
            )
            logprobs_supported = False
            continue

        last_error = e
        logger.warning(f"HTTP error from {endpoint.name}: {e}")
        if attempt < self.settings.ai_max_retries - 1:
            continue