Skip to content

Journal

·5 min read

Ollama insights see aggregates. Never a visitor.

Weekly write-ups run on rollups — counts, ratios, deltas. Point the provider at Ollama and nothing leaves the machine. Event-level rows never enter the prompt.

OllamaInsights

A compact home workstation at night with a single monitor
The brief can be written on the same machine that stored the rollups.

Asking a hosted model to “summarise this week” is easy and usually reckless: people paste event exports into a chat. Vector 5 never does that. Insights are generated from rollups only — page counts, source shares, actor splits, week-over-week deltas. There is no visitor hash, no IP, no user agent string in the prompt.

Ollama is the default local provider. llama3.2 on your own box produces the same shape of brief as OpenAI, Anthropic, Gemini, or any OpenAI-compatible endpoint — without the week’s traffic leaving the server. Turn insights off and no model is called at all.

What the model is allowed to see

  • Counts and ratios for the selected window.
  • Actor mix: human, human_via_ai, ai_crawler, ai_agent, bot.
  • Top pages and sources, including crawl-to-referral rates.
  • Anomaly flags: a crawler spike, an assistant-referral surge.

A typical brief might say that ChatGPT now sends more humans than Reddit, or that GPTBot fetched /pricing hundreds of times for two referred visitors. Those sentences are computed from aggregates you can already see in the dashboard. The model is a writer, not a second database.

If the prompt cannot be rebuilt from the public charts, it does not belong in the prompt.