@silicajs/assistant

Optional AI assistant for Silica knowledge sites.

The assistant answers reader questions from the site's original markdown files, cites the source pages it used, and supports multi-turn follow-ups. It is built on core-ai for provider-agnostic model access and uses an in-process, read-only shell sandbox (just-bash) for markdown exploration (ls, grep, cat, …).

Usage

Install alongside the provider package for your model:

npm install @silicajs/assistant @core-ai/openai

Enable it in silica.config.ts:

assistant: {
  provider: "openai",
  model: "gpt-5-mini",
},

and set the provider API key (e.g. OPENAI_API_KEY) plus a server-only SILICA_ASSISTANT_SECRET in your environment. The Silica CLI generates the /api/assistant route and passes the assistant UI to the active theme.

For providers without a Silica preset, configure the core-ai package and factory explicitly:

assistant: {
  provider: {
    package: "@acme/core-ai-provider",
    factory: "createAcme",
    env: { baseURL: "ACME_BASE_URL" },
    secrets: { apiKey: "ACME_API_KEY" },
    options: { region: "eu" },
  },
  model: "acme-chat",
},

Generated assistant routes include a 10 requests/minute rate limit keyed by x-forwarded-for. In silica.config.ts, set assistant.rateLimit.trustedProxyHeaders if your deployment proxy uses a different client-IP header, and only include headers your proxy sets or overwrites.

Entry points

| Export | Contents | | ---------------------------- | ---------------------------------------------------------------------------- | | @silicajs/assistant | Shared types (citations, transcript, stream events) | | @silicajs/assistant/ui | Client components: AssistantProvider, AssistantTrigger, AssistantPanel | | @silicajs/assistant/server | Framework-agnostic runtime: handler, agent loop, sandbox, citations | | @silicajs/assistant/next | Route glue for the generated Silica Next.js app |

See the Silica docs for the full feature documentation.