-
Engineers place the model in a replaceable layer.
-
The company's knowledge and processes remain in its own infrastructure and survive any vendor change. ### A gateway instead of direct calls
-
All model requests go through a single point—in our case, the LLM & Security Gateway.
-
The gateway routes requests between OpenAI, Anthropic Claude, Google Gemini, GigaChat, YandexGPT, and open models such as Qwen and Llama; masks personal data; and logs usage and quality.
-
To switch to a new model, the team changes the gateway configuration.
-
The process code remains unchanged. ### Tools through MCP
-
The model's access to 1C, PIM, ERP, and databases is defined once through MCP.
-
Any model that supports the protocol gets the same tools.
-
The company's knowledge is stored separately in a RAG index, so it does not need to be moved when the model changes. ### Your own test suite
-
A company needs its own reference set: 50–200 real tasks from its process with correct answers.
-
When a new model is released, the team runs it against this set and sees within a day whether the metric has changed.
-
The vendor's benchmark table shows an average result on someone else's tasks. ### Long outputs require oversight
-
One million output tokens are useful where large volumes of repetitive text are needed: product descriptions and attributes for Pimcore or Akeneo, code migration, and documentation. In our PIM projects—product card text, taxonomy tagging, and filter facets—validation takes the most effort.
-
Before publication, each generated product card is checked against the attribute schema and category tree.
-
Without this validation, a long response becomes a mountain of text that no one reads and no one is accountable for.
