Skip to content
parelyxDevelopers

Keep your SDK. Change two values.

Parelyx exposes an OpenAI-compatible interface, so existing clients, agent frameworks and RAG pipelines integrate by changing the base URL and key. Your endpoint and credentials are issued during onboarding.

  1. 01
    Request access

    Tell us about your workloads. Once onboarded, you receive a Parelyx endpoint and API key.

  2. 02
    Point your client

    Set the base URL and key in your existing OpenAI-compatible client.

  3. 03
    Baseline, then enable

    Run with optimisations off to establish a baseline, then enable strategies where the evidence supports them.

API integration

Reference pattern

The highlighted lines are the integration. Everything else is your existing code. The endpoint and key below are placeholders read from your environment.

1import os2from openai import OpenAI34client = OpenAI(5    base_url=os.environ["PARELYX_BASE_URL"],6    api_key=os.environ["PARELYX_API_KEY"],7)89response = client.chat.completions.create(10    model="your-model",11    messages=[{"role": "user", "content": "Summarise the refund policy."}],12)13print(response.choices[0].message.content)
app.pyOpenAI-compatible client
1# Issued during onboarding2PARELYX_BASE_URL=<your Parelyx endpoint>3PARELYX_API_KEY=<your Parelyx key>
.env

The complete API reference is provided during onboarding.

Requests are unchanged
Send the same request body your application already sends. Parelyx evaluates it; it does not require new fields.
Responses are unchanged
Your application receives a standard response, whether it came from a model or from semantic reuse.
The model field
Identifies the model or route configured for your deployment. Selection rules map it to an endpoint.
Every request has a decision record
What was decided, why, and the observed outcome — available in observability, not in your response path.

Hosted model providers and self-hosted models that expose an OpenAI-compatible interface are registered for your deployment during onboarding. Selection policies then determine which endpoint serves each request.

  • Hosted model APIs
  • Self-hosted open-weight models
  • Multiple providers per deployment

Every optimisation is configurable and can be switched off. Explicit configuration takes precedence over policy.

Semantic response reuse
Enable or disable reuse, and set how closely a request must match before a validated response is returned.
Prompt optimisation
Enable or disable conservative removal of redundant context before execution.
Model & provider selection
Define how requests are mapped to the models and endpoints registered for your deployment.

Each request produces a decision record and outcome — tokens, latency, selected strategy and endpoint, and estimated cost. See Observability for the measures that matter and what each one tells you.

Parelyx runs in your environment, alongside your applications and models. Deployment topology, data retention and access controls are agreed during onboarding.

Responses are slower than calling the model directly
Check the decision record for the request in observability. If an optimisation was applied that did not pay off, disable it for that workload. If the overhead is in the decision itself, report it to us — bounded decision overhead is a design requirement.
A reused response didn't fit the request
Raise the similarity threshold for that workload, or disable reuse for workloads whose answers are personalised or time-sensitive.
A request went to a different model than expected
Review the selection rules configured for your deployment, and check the selected endpoint in the request's decision record.
Requests are failing
Parelyx is designed to fail open, so faults inside Parelyx should not fail requests. Errors returned from the model endpoint itself are passed back to your application. Contact engineering with the time window and workload if you see otherwise.
Ready to integrate?
Endpoints, keys and the full API reference are issued during onboarding.
Request access