llm_id2llm_type ignores the factory part of "name@factory", causing wrong model_type for models that exist in multiple factories

Description

TenantLLMService.llm_id2llm_type (api/db/services/tenant_llm_service.py) can return the wrong model_type for a chat model, causing LookupError: Model(<name>@<factory>) not authorized in dialog_service.async_chat, even though the model is correctly configured in tenant_llm with model_type='chat'.

Root Cause

def llm_id2llm_type(llm_id: str) -> str | None:
    from api.db.services.llm_service import LLMService

    llm_id, *_ = TenantLLMService.split_model_name_and_factory(llm_id)
    llm_factories = settings.FACTORY_LLM_INFOS
    for llm_factory in llm_factories:
        for llm in llm_factory["llm"]:
            if llm_id == llm["llm_name"]:
                return llm["model_type"].split(",")[-1]
    ...

split_model_name_and_factory correctly splits llm_id (format "<name>@<factory>") into (name, factory), but the factory part is discarded (, *_). The subsequent loop iterates over all factories in settings.FACTORY_LLM_INFOS and returns the model_type of the first llm_name match, regardless of which factory it belongs to.

Concrete Example

gemini-2.5-flash exists in conf/llm_factories.json under multiple factories, e.g.:

• An earlier factory entry (appears first in the JSON array) with "model_type": "image2text"
• The "Google Cloud" (Vertex AI) factory entry with "model_type": "chat"

For llm_id = "gemini-2.5-flash@Google Cloud":

1. split_model_name_and_factory returns ("gemini-2.5-flash", "Google Cloud"), but only "gemini-2.5-flash" is kept.
2. The loop finds the first llm_name == "gemini-2.5-flash" match — which belongs to a different factory — and returns its model_type, "image2text".
3. Back in dialog_service.async_chat:

   llm_type = TenantLLMService.llm_id2llm_type(dialog.llm_id)
   if llm_type == "image2text":
       llm_model_config = TenantLLMService.get_model_config(dialog.tenant_id, LLMType.IMAGE2TEXT, dialog.llm_id)
   else:
       llm_model_config = TenantLLMService.get_model_config(dialog.tenant_id, LLMType.CHAT, dialog.llm_id)

   Since llm_type == "image2text", it queries tenant_llm for model_type='image2text' for this tenant — which does not exist (the tenant only configured model_type='chat' for gemini-2.5-flash@Google Cloud), raising:

   LookupError: Model(gemini-2.5-flash@Google Cloud) not authorized
   

Why a DB workaround is not possible

Adding a second tenant_llm row with the same tenant_id/llm_factory/llm_name but model_type='image2text' is blocked by:

UNIQUE KEY `uk_tenant_llm` (`tenant_id`,`llm_factory`,`llm_name`)

Suggested Fix

llm_id2llm_type should use the factory information from split_model_name_and_factory to only match llm_name within the specified factory (falling back to the current name-only behavior if no factory suffix is present):

def llm_id2llm_type(llm_id: str) -> str | None:
    from api.db.services.llm_service import LLMService

    llm_name, llm_factory_name = TenantLLMService.split_model_name_and_factory(llm_id)
    llm_factories = settings.FACTORY_LLM_INFOS
    for llm_factory in llm_factories:
        if llm_factory_name and llm_factory["name"] != llm_factory_name:
            continue
        for llm in llm_factory["llm"]:
            if llm_name == llm["llm_name"]:
                return llm["model_type"].split(",")[-1]

    for llm in LLMService.query(llm_name=llm_name):
        return llm.model_type

    llm = TenantLLMService.get_or_none(llm_name=llm_name)
    if llm:
        return llm.model_type
    for llm in TenantLLMService.query(llm_name=llm_name):
        return llm.model_type
    return None

Environment

• RAGFlow version: v0.25.6
• LLM factory: Google Cloud (Vertex AI), model: gemini-2.5-flash
• Deployment: Docker, infiniflow/ragflow:v0.25.6 image