iNew developer tools are available in the Autumn release.×

Introduction

iLooking for the Inkstone workspace? Open Canvas from your account menu.

The Inkstone API can support research, writing, analysis, and automation. Choose from a family of language models with different capabilities and operating profiles, or adapt a model to your domain with guided tuning.

Resources

iProtecting customer data is a core part of our service. API inputs and outputs are not used to train shared models. Review the data stewardship guide.

Key concepts

Language models

Inkstone language models are trained to understand and produce natural language. A request supplies instructions and context; the model returns a useful completion that follows those directions. Different model families balance depth, response time, and cost, making them suitable for tasks from classification to long-form synthesis.

Clear instructions and concrete examples improve reliability. Start with the text generation guide, then use our prompt patterns to refine behavior for your application.

Workrooms

Workrooms are durable agents built around a chosen model, a set of instructions, and optional tools. They can preserve relevant context across a session, run approved calculations, and retrieve information from connected documents. Learn how to design one in the workrooms overview.

Embeddings

An embedding is a numerical representation of meaning. Related text appears close together in the resulting vector space, enabling semantic search, recommendations, anomaly detection, clustering, and classification. Inkstone embedding models accept a string or collection of strings and return compact vectors for downstream systems.

Tokens

Models process text in small pieces called tokens. A token may be a short word, part of a word, punctuation, or spacing. In English, one token is roughly four characters, though the exact count varies with language and content. Use the token counter to estimate a request before sending it.

Each model has a context window shared by the input and generated response. Requests that exceed that limit must be shortened or split into stages. Current context limits and output allowances are listed in the model catalog.

Guides

Start with the guide closest to what you want to build.

Was this page useful?