Cerebras/cerebras-dataiku-plugin

Cerebras/cerebras-dataiku-plugin

Language: Python

License: Apache-2.0

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-10-05T19:17:02Z

Pushed: 2026-10-12T16:54:34Z

Default branch: main

Fork: no

Archived: no

README:

Cerebras Inference for Dataiku

Bring Cerebras Inference models into the Dataiku LLM Mesh as a native LLM connection.

Once configured, Cerebras models are available everywhere Dataiku offers an LLM — Prompt Studios, LLM recipes, Dataiku Answers, agents, knowledge banks, and the Python APIs — with no code changes to your existing projects.

Features

| | | | --- | --- | | Chat completions | Full support for single and multi-turn conversations | | Streaming | Token-by-token responses in Prompt Studios and Answers | | Tool calling | Multi-turn tool use, including parallel tool calls, in both streaming and non-streaming modes | | Structured outputs | JSON object and JSON schema response formats | | Reasoning | Configurable reasoning effort, with reasoning traces captured for inspection | | Vision | Inline PNG and JPEG prompts on vision-capable models | | Usage reporting | Prompt and completion token counts surfaced to the LLM Mesh | | Credentials | One admin-managed key for everyone, or an individual key per user |

This connection provides chat completions.

Requirements

• Dataiku 14.1 or later
• A Cerebras API key from cloud.cerebras.ai

No code environment to build: the plugin's only dependency is requests, which is already present in the Dataiku builtin Python environment.

Outbound HTTPS access from the Dataiku backend to api.cerebras.ai.

Installation

1. Install the plugin

In Dataiku, go to Administration > Plugins > Add plugin and either upload the plugin zip or add it from the store. There is no code environment to build, so the plugin is usable as soon as it is installed.

2. Add your API key

Choose one of the two credential models — see [Credentials](#credentials) for the trade-off.

*Shared key (simplest).* Go to Administration > Plugins > Cerebras Inference > Settings > Shared Cerebras API key and create a preset. Paste your Cerebras API key and save. The key is stored encrypted; end users of the connection never see it.

*Individual keys.* Create an empty Individual Cerebras API key preset instead, and grant it to the relevant groups. Each user then goes to Profile and settings > Credentials > Plugins and enters their own key.

In both cases, grant preset access only to the users or groups who should have it.

3. Create the connection

Go to Administration > Connections > New connection > Custom LLM > Cerebras Inference.

| Field | What to enter | | --- | --- | | Credential mode | Shared key (admin-managed) or Individual key (per user) | | Shared / Individual Cerebras API key | Select the preset from step 2 | | Model | A Cerebras model ID, for example gpt-oss-120b | | Base URL | Leave as-is unless you use a dedicated endpoint |

Only the preset field matching your chosen mode is shown.

Name the connection, save, and grant access to the relevant user groups.

4. Verify it works

Open a Prompt Studio in any project, select your new Cerebras connection, and send a test prompt. You should see a response stream back.

Configuration

All settings live on the connection, so you can create several connections against the same API key — one per model, or per team.

| Setting | Default | Description | | --- | --- | --- | | Credential mode | Shared | Where the key comes from. See [Credentials](#credentials) | | Shared Cerebras API key | — | The admin-managed preset. Shown in Shared mode | | Individual Cerebras API key | — | The per-user preset. Shown in Individual mode | | Base URL | https://api.cerebras.ai/v1 | The Cerebras API endpoint | | Model | gpt-oss-120b | The model ID to serve. See [Choosing a model](#choosing-a-model) | | Force image input support | Off | Enables image prompts for a vision-capable model the plugin does not recognize yet. Well-known vision models are detected automatically. Does not relax the PNG/JPEG or inline-only rules | | Reasoning effort | Model default | none, low, medium or high. Controls how much the model reasons before answering, on models that support it | | Include reasoning in the answer | Off | When on, reasoning is returned with the answer inside ` tags. When off, it is recorded on the query trace only | | **Request timeout (s)** | 300 | Maximum time for a single request | | **Max retries** | 3 | Automatic retries for rate limits and transient errors, using exponential backoff | | **Extra payload fields** | — | Additional Cerebras API parameters to send with every request, such as prompt_cache_key or parallel_tool_calls` |

Credentials

Pick the mode that matches how you want Cerebras usage attributed.

| | Shared key | Individual key | | --- | --- | --- | | Who supplies the key | An administrator, once | Each user, for themselves | | Where it is entered | Plugin settings preset | Profile and settings > Credentials > Plugins | | Cerebras billing | All usage on one key | Attributed per user | | Cerebras rate limits | Shared by all users | Each user has their own | | Users need a Cerebras account | No | Yes |

Shared is the simpler operational model and the default: one key to rotate, and users need nothing from Cerebras. Use separate presets per team if you want their usage billed or rate-limited apart.

Individual is the right choice when you need per-user attribution, when one user's traffic must not consume everyone else's rate limit, or when policy forbids a shared service credential. The administrator creates the preset empty; it stays unusable until each user fills in their own key.

A single connection uses one mode. To offer both, create two connections.

Choosing a model

Enter the model ID exactly as Cerebras publishes it, for example gpt-oss-120b. Because this is a free-text field, new Cerebras models and custom weights on dedicated endpoints work immediately, with no plugin update required.

Browse the current models, context windows and pricing in the Cerebras model catalog.

To list the models available to a specific key:

curl https://api.cerebras.ai/v1/models \
-H "Authorization: Bearer $CEREBRAS_API_KEY"

To serve several models at...