Skip to main content
The Sphinx Responses API augments your agent’s inference requests with learned context. Sphinx can inject context over multiple rounds and returns a response in OpenAI’s Responses format.

Before you start

  • Ask Sphinx to enable API access for your organization and confirm the workspace and model configuration for your integration.
  • Obtain a Sphinx API key for that setup.
  • Use an agent that sends Responses API requests and lets you change the endpoint and authentication.

Configure your agent

Replace your agent’s inference endpoint and credentials with the Sphinx values.

Supported OpenAI models

Set the request’s model field to one of these exact IDs, without an openai/ prefix: These are the OpenAI model IDs supported by the Responses API. Availability also depends on your deployment’s provider configuration; confirm access with Sphinx. Unlisted model names, including unlisted aliases or dated snapshots, are rejected with model_not_found.
Use a Sphinx API key for this endpoint. The MCP endpoint is https://api.prod.sphinx.ai/mcp and has a separate client setup flow.

Send a first request

Set SPHINX_API_KEY to your Sphinx key and SPHINX_MODEL to the enabled model name in your environment, then send a minimal request:
Choose a question covered by your workspace’s knowledge. Check that the request succeeds, your agent can read the Responses-format result, and the answer uses the intended business definitions.
Keep the key in your environment or secret store. Do not embed it in source code, publish it in configuration, or include it in screenshots and logs.

Integrate with an existing agent

Keep your agent’s task instructions and source-access tools in place. If it needs local documents, specify their location in its system instructions. If it needs a database, provide the appropriate query tools and access separately. Test the request options your agent uses, including streaming or tool interactions, before switching a production workflow. Contact Sphinx about any client-specific compatibility requirements.

Evaluate the integration

Compare the same agent and tasks with and without Sphinx. Hold the model, source access, and success criteria constant, then compare answer quality, cost, and latency. Include new tasks that were not used to develop the learned context. For incorrect answers, inspect the relevant knowledge, correct the issue, and rerun your evaluation. Use Benchmarks (beta) for recurring checks; contact Sphinx to activate it.