Tarseek API · Pre-production

Responses API

POST /v1/responses provides stateless text and client-function generation through the same Tarseek permissions, project limits and BYOK credentials as Chat Completions.

from openai import OpenAI
import os
client = OpenAI(api_key=os.environ["TARSEEK_API_KEY"], base_url="https://tarseek.com/v1")
with client.responses.stream(model=os.environ["TARSEEK_MODEL_ID"], input="Hello", max_output_tokens=512, store=False) as stream:
    for event in stream:
        if event.type == "response.output_text.delta":
            print(event.delta, end="", flush=True)
    response = stream.get_final_response()

Supported requests

String input or a full history of user, assistant, system and developer messages; text, image URLs/data URLs and Tarseek file IDs; instructions; max_output_tokens; temperature/top_p; client function tools and function_call_output; parallel tool control; text.format JSON schemas; low/medium/high reasoning effort where the selected provider supports it. Project policies and response-cache controls also apply.

Stateless contract

Include the full conversation on every call. store=true, previous_response_id, background execution, conversation references, server tools and encrypted reasoning items are rejected. No response retrieval, deletion or cancellation endpoint is provided. Opt-in response caching is separate from conversation storage and follows the project-controls retention rules.

Streaming and completion

Text and function arguments stream incrementally with stable output indexes, item IDs and sequence numbers. Plain provider reasoning is included in the final reasoning item; encrypted chains and summary generation are not supported. Token limits produce an incomplete response with max_output_tokens as the reason. Failed streams emit an error and response.failed, never response.completed. Accumulated stream output is limited to 4 MiB; client tool calls are limited to 128. Usage comes from provider counters and cache replays report zero billable usage.