ai-studio

Create an OpenAI client with base_url https://ai.api.cloud.yandex.net/v1, your api_key, and project set to folder_id, then call client.responses.create with a model URI like gpt://{folder_id}/qwen3-235b-a22b-fp8. Credentials can come from a .env file.

Use qwen3-235b-a22b-fp8 for general text and tools, qwen3.6-35b-a3b for image input, deepseek-v4-flash for inputs up to 1M tokens, and aliceai-image-art-3.0 for direct image generation. The models reference lists exact URIs and context limits.

Define a Pydantic BaseModel and pass it as text_format to client.responses.parse. The parsed result is available as response.output_parsed, avoiding manual JSON extraction.

Yes. Upload files with purpose assistants, add them to a vector store with create_and_poll, then query using the file_search tool in a Responses API call. Citations appear as file_citation annotations on the response.

Never execute generated code locally with exec. Use the hosted Code Interpreter tool with an automatic or explicit container, which runs code in a sandbox and returns files as container_file_citation annotations.

Use the Agents SDK only when you explicitly need Runner, handoffs, or multi-agent behavior. Ordinary tool-using agents should use direct Responses API patterns with function tools, web search, or hosted MCP servers.