This is the documentation for Dreamspace, the prompt diagramming tool, a node-based infinite canvas for running prompts and inspecting the outputs of LLMs.

How It Works

You use the application by adding nodes to the canvas; some represent prompts and all their parameters, others represent outputs like images, chat messages, and text. From any node, you can chain to create a new prompt with a different model. You can also add messages to a chat thread, or upload images for vision models.

Directed arrows connect prompts to outputs, outputs to chained prompts, and messages within a chat thread. All of your latent space explorations are saved in the current "space" and are private by default. You can create a new space at any time, and navigate between them in the sidebar or "My Spaces" preview page.

Getting Started

After signing up for the application, you will be redirected to our onboarding flow. Here you can set up your first model connections, and select a plan. Dreamspace is a paid digital subscription product.

Your subscription provides secure cloud storage to automatically save all of your prompts, outputs like images, text, and chat histories, ensuring you never lose any important exploration or output. This persistent storage allows you to seamlessly continue your work across sessions and devices, and revisit experiments over time.

The Pro subscription is best for individuals and teams looking for consulting from our prompt engineers, early access to premium tools, and highest priority for application feedback and support.

Connect to Models

The application uses an API key that you provide to connect to supported OpenAI and Replicate-hosted models. You can also connect to models hosted elsewhere by providing an API endpoint and valid configuration.

OpenAI is the recommended model provider to use with Dreamspace based on availability and latency over thousands of sessions. Compatible models will be added as they become available.

Replicate is an alternative model provider with many more models such as Stable Diffusion and LLaMA. However, runtimes can vary dramatically between models and sessions due to Replicate architecture. Compatible models can be added by request.

Anthropic support was most recently added for Claude 3 Opus and Sonnet. Compatible models can be added by request.

Create a Space

From the Home page you can create a new space. You can also view a preview of your saved spaces and navigate elsewhere within the application.

From within an existing space, pressing the Dreamspace logo in the upper-left opens the navigation drawer. From here, you can create a new space, view a list of your spaces, search across them, and navigate to other pages.

Run Prompts

By pressing the "Add Prompt" button in the upper toolbar, your cursor will move into prompt placing mode. New spaces are in placing mode by default.

Before a prompt is run, you can change everything about it: the model, settings, and prompt text.

Once you run a prompt node, the model can no longer change for that node or its children without chaining. However, the configuration settings and prompt text can change. To preserve output ancestry, it is not recommended to change the prompt text of nodes with children. Instead, chain to create a copy which can be modified.

Because the example's model follows a "chat" schema, you can also manually add messages to the conversation thread. This example shows three GPT generated responses from the original prompt, and then some manually written instruction nodes, forking to introduce additional context on the formatting of responses. The result of this exploration is a prompt that can be run to more closely match the format the user provided.

Compare Outputs

The infinite canvas is a conventient medium for sampling latent space many times with different prompts and settings. Click on any node to view its configuration. This example shows how a prompt provided to dall-e-3 is modified and returned along with the response. The output node can then be run again or chained with the output configuration as input for the next.

Chain Them Together

An important concept in Dreamspace is the ability to chain the output from one prompt into the input of a new prompt. This is useful when comparing variations between prompt texts and models, as well as when the output needs additional processing. The example below shows a vision node, one of its outputs being chained into a new image node, and one final chain into a node for another image model.

Frequently Asked Questions

Why does Dreamspace cost $10/month?

Dreamspace has a free demo with full functionality except the spaces you create cannot be saved. There is a paid subscription so that all of your prompts and outputs like images and chat threads can be saved and accessible across devices and sessions. Your subscription goes a long way to making sure we can continue to release high-quality prompt engineering tools.

Why do I have to provide my own API key?

Dreamspace is a infinite canvas user-interface for interacting with AI models. We do not host or train our own models (yet), and rely on access to services that do. By providing your own API keys, content moderation and usage based billing is managed by model providers.

What's next?

✦ New models and providers

✦ Automatic layouts

✦ Recommended prompts

✦ Public spaces

✦ and much, much more