Suggestions

The platform suggests capability packages for a prompt, a task or a project.

Built-in rules (default)

Suggestions come from technologies, keywords and categories found in the text. No text leaves the server.

• POST /api/v1/orgs/:orgId/registry/suggest takes any mix of text, projectId and taskId. A project adds its name, description, knowledge and the stack its last readiness check found; a task adds its title, prompt and background.
• Each suggestion comes with its reasons ("Works with Playwright", "Testing & QA") and whether it is already installed for that task, project, person or organization.
• Shared technologies weigh most, then the package's own keywords, then categories. A shared category alone is never enough. Curated packages come first.
• The dashboard shows suggestions in Capabilities → Suggested and as you write a new task.

Semantic suggestions (optional)

With an embeddings API configured, packages close in meaning to the description are suggested too, even when they share no words with it.

EMBEDDINGS_URL=https://api.openai.com/v1   # any OpenAI-compatible API, e.g. Ollama: http://localhost:11434/v1
EMBEDDINGS_API_KEY=…                       # when the API needs one
EMBEDDINGS_MODEL=text-embedding-3-small
EMBEDDINGS_MIN_SIMILARITY=0.4              # depends on the model

• What is embedded: each package's name, description, tags and the start of its readme, in the background at startup and when a listing changes. Server → Marketplace can start a full run (POST /admin/registry/embed). Changing the model embeds everything again.
• Ranking: closeness in meaning adds to the rules' score. Curated packages still come first.
• Failure: when the API fails, suggestions fall back to the rules.
• Scale: without Atlas, the 5,000 most used public packages are compared in memory. With REGISTRY_SEARCH=atlas, create a vector index named capability_embeddings on embedding.vector.