Assistant

How the Cueset Labs assistant works: the self-hosted language model behind it, how it retrieves approved site knowledge, and the limits it runs under.

The Cueset assistant runs on a private Ollama instance and answers questions grounded on an approved knowledge base. This page documents the current model, retrieval configuration, and operational limits.

What it can answer

The assistant is powered by an open-weight language model running on a private server. It is not connected to the internet and has no access to real-time information. It answers general questions from its training, and questions about Cueset Labs from approved reference material supplied with each question, and it says so when that material doesn't cover what was asked. It cannot browse the web, send emails, submit forms, access accounts, or take actions on your behalf.

How an answer is produced

1. Limits: each question counts against a per-visitor and per-conversation rate limit, then waits for a generation slot. The server runs on CPU, so questions queue rather than run side by side.
2. Retrieval: the question is matched against the approved knowledge base by meaning and by shared keywords, and falls back to keywords alone when embeddings are unavailable.
3. Prompt: the system prompt, the matched documents and the page you are reading are combined into one set of instructions for the model.
4. Streaming: the answer streams back as it is written, together with links to the site pages it drew on.
5. History: the server keeps no conversation state. Your browser resends the recent turns with each question, trimmed to a fixed length.

What this page shows

• Model: the model the assistant runs and the version of the knowledge it draws on.
• Retrieval: how it finds approved material for a question, and how many documents it uses.
• Context and limits: the input, history and output limits every request runs under.
• Knowledge base: a searchable list of the documents the assistant can cite.
• Eval harness: deterministic policy checks you can run against its answers.

Privacy

The API binds to loopback behind a TLS reverse proxy. Ollama is never exposed publicly. Prompts, messages, and conversation IDs are logged with request metadata only. No message content is written to persistent logs. The server holds no conversation state; the client resends history each turn.