eInk Label GenAI Templates

A companion plugin for @rhizomatics/signalk-einklabel-plugin that generates an image to paint on shelf label entirely from an LLM prompt, as an alternative to a hand-authored SVG template.

By default it uses OpenRouter, however it can use any of the Vercel AI SDK providers directly, including local LLMs like Ollama.

Working Example

This plugin comes with a marine weather forecast example prompt. When installed, and SignalK server restarted, you should see this appear as a template to choose in the main eInk Label plugin.

Expanded Example

Forecast

Summarize the maritime weather outlook for the next 24 hours, appropriate for the sailor of a 13.6m Sailing craft positioned at latitude 40.689247, longitude -74.044502

Target Guidance

The following has been automatically added to the end of the standard prompt, with information specific to which label is being targetted.

Generate an SVG image for the result suitable for a ZhunyCo 4.13" eInk Electronic Shelf Label display with pixel height 240 and width 416. Available colours are black (#000000), white (#FFFFFF), red (#FF0000), yellow (#FFFF00). Fonts available are serif, sans-serif, monospace.

Use graphics to make it easier to read and use the space better, and use colour and layout to draw attention to safety critical information. Include in minimal size font the source of the information and generation timestamp.

CAUTION

Be careful with GenAI generated content, it will often be very good but may be prone to hallucinations or omissions.

Testing on your server

If you don't already have an AI account, get a free API key from OpenRouter - set the model to one of the specific free models, or openrouter/free, OpenRouter's free-tier routing alias, although note you will get a random model every time it is called, some of which are specialized (e.g. "content safety") and will not generate an image. Free models may not lookup web pages or APIs, so guess things like weather, and the image quality will likely be much poorer than using a mainstream LLM.

Use the esl-cli CLI to see generated prompts or create test PNGs.

Customizing Prompts

Prompts have the same template syntax as the SVG images, so the weather forecast will give the current GPS position, and the length and type of boat.

All prompts also get a generated appendix to the prompt that gives details about the eInk label, including manufacturer, name, pixel height and width, colour range, and the description provided in the main plugin.

Installation

Install both plugins from the SignalK AppStore (or npm install @rhizomatics/signalk-einklabel-plugin @rhizomatics/signalk-einklabel-genai-plugin), then install whichever LLM provider package(s) you'll actually use - these are optionalDependencies, not installed automatically, since most people will likely only ever need one. openrouter is the one exception: @openrouter/ai-sdk-provider is a regular dependency, so it's already installed and needs no extra step, as is the base ai package:

npm install ai @ai-sdk/openai          # llmProvider: "openai"
npm install ai @ai-sdk/anthropic       # llmProvider: "anthropic"
npm install ai @ai-sdk/google          # llmProvider: "google"
npm install ai @ai-sdk/xai             # llmProvider: "xai"
npm install ai @ai-sdk/deepseek        # llmProvider: "deepseek"
npm install ai @ai-sdk/moonshotai      # llmProvider: "moonshotai"
npm install ai ai-sdk-ollama           # llmProvider: "ollama"
npm install ai @ai-sdk/openai-compatible  # llmProvider: "local" (LM Studio, vLLM, ...)

If the package for your configured provider isn't installed, a GenAI-backed device falls back to the core plugin's warning template with a log message naming exactly which package to npm install.

Local LLMs

Many Raspberry Pi type tiny computers either have an onboard NPU for AI work, or an add-on 'hat', like the Raspberry Pi AI Kit.

Combined with a distilled model, for example Llama 3.2 3B, Phi-4 Mini 3.8B, or Qwen3.5-4B, it is possible to have something reasonable in 8Gb, and the eInk use case is ideal, since it runs in the background a couple of times a day, so slow token generation isn't an issue.

Configuration

• llmProvider - openai, anthropic, google, xai, deepseek, moonshotai, openrouter, ollama, or local (any other OpenAI-compatible server). The config screen's dropdown only lists providers whose package is actually installed (openrouter always is) - install another one's optional package and reopen the page to add it.
• llmApiKey - not needed for ollama/local unless your server itself checks one
• llmModel - provider-specific model id, e.g. gpt-4o, claude-sonnet-4-5, gemini-2.5-flash, grok-4, or an Ollama/local model tag
• llmBaseUrl - required for local; optional for ollama (defaults to http://localhost:11434/v1)
• llmTimeoutSeconds/llmRetries - how long to wait, and how many attempts, before giving up (the core plugin then shows its fallback warning)
• promptsDir - where to look for your own .md prompt files, same convention as the core plugin's templatesDir (empty for the default, a relative path resolves against ~/.signalk, absolute used as-is)
• testConnection - not a persistent setting, just a one-shot checkbox: check it and save to send one test message to the LLM settings above as soon as the plugin (re)starts. The result (success, or the error) shows up on this plugin's own entry on the Plugin Config page - the checkbox clears itself afterwards, so it won't refire on the next ordinary restart.

Writing Prompts

A prompt (.md file) is just prose, with the same {...} binding syntax as a template's <desc> in the core plugin (see its README, "Template Source Specification") - a bare path like {design.length.overall}, or the full source=...,path=...,format=... grammar, resolve exactly the way they would in a template. One more source is available here, source=label, giving facts about the physical label the LLM needs to design for:

• path=width/path=height - exact pixel size the returned SVG must match
• path=colours - the label's colour palette (add format=csv for a plain comma-separated list)
• path=fonts - the three font-family keywords the renderer can actually display (also supports format=csv)
• path=description - the device's own configured Location/description field
• path=manufacturer/path=label - the panel's real-world brand and its own size label (e.g. 3.7")
• path=position - the vessel's current GPS position, pre-formatted as decimal degrees with hemisphere letters

A placeholder that doesn't resolve to any value becomes ??? in the prompt text - a visible gap the LLM can reason around, rather than a silent, misleading blank. Add default=<value> to any binding to use that value instead when the data is missing.

Every prompt also has a bundled .assets/target_guidance.md fragment appended automatically, telling the model the exact size/colour/font constraints and that it must respond with nothing but a single self-contained <svg>...</svg> document. Override it the same way you'd override any bundled prompt - put your own .assets/target_guidance.md in your promptsDir.

Testing Prompts from the CLI

This plugin contributes prompt/generate subcommands to the core plugin's esl-cli, loaded via -r:

esl-cli -r @rhizomatics/signalk-einklabel-genai-plugin/cli prompt forecast -e examples
esl-cli -r @rhizomatics/signalk-einklabel-genai-plugin/cli generate forecast -e examples -o forecast.png --llm-provider ollama --llm-model llama3.2

• prompt [name] - resolve a prompt's placeholders against a live SignalK server (or example data) and print the exact text an LLM would be sent, with no LLM call
• generate [name] - call the configured LLM and write the rendered SVG's PNG, optionally saving the raw SVG too with --save-svg

The optional [name] argument names a prompt the same way the config UI's picker does - a full path also works. Use --help on either for the full option list.

Architecture

Why 2 plugins?

This is deliberately a separate plugin, not a feature of the core one: the core plugin has nothing to do with any LLM SDK or API key, so installing it never pulls in GenAI dependencies unless you also explicitly install and enable this one - useful for anyone on a small/constrained install, or who'd simply rather not have any GenAI code in their install at all.

How it fits together

This plugin registers a TemplateProvider with the core plugin (the same extension mechanism a new vendor's hardware driver uses - see the core plugin's README, "Extending"). Once both plugins are installed and enabled:

• Its prompts show up as ordinary entries in the core plugin's own "Template" picker, suffixed (GenAI) - pick one exactly like you'd pick a .svg file. No separate "render mode" setting.
• All LLM provider/model/API-key configuration lives in this plugin's own config screen - the core plugin knows nothing about any of that.
• If the LLM call fails (network/API error, exhausted retries) or its response isn't a renderable SVG, the core plugin pushes its own generic fallback-warning template instead of leaving the previous, possibly now-wrong, content on screen - the same safety net a broken hand-authored template gets. The next scheduled repaint retries automatically.

Since each repaint may call a (usually paid) LLM API, a GenAI-backed device should use Repaint Trigger interval, not subscription.

Adding another LLM provider

The provider list above is resolveModel's (src/llmGateway.ts) built-in switch, but you're not limited to it: call registerLlmProvider(name, factory) (exported from this plugin, mirroring the core plugin's own extension points) from your own SignalK plugin's start() to add support for a provider this plugin doesn't bundle, without forking it. A registered name takes priority over the built-in switch, so it can even replace one of the names above.

import genai from "@rhizomatics/signalk-einklabel-genai-plugin";
genai.registerLlmProvider("my-provider", (settings) => myOwnModelFactory(settings));

Declare this package as a regular npm dependency (not a peerDependency - see the core plugin's own README for why) plus "signalk": { "requires": ["@rhizomatics/signalk-einklabel-genai-plugin"] } in your extension's package.json.