🔎 Reference appBuilt in
Web search
Results shaped for an assistant with limited memory.
What you'll learn
- Why every result is short and says how to continue.
- How untrusted addresses are handled.
- How to use search from your own skill.
Search and page reading are built into Tallgeese, so you don't add them. You use them from your skills and specialists, as the Financial Advisor does. They're also a good example of how to write tools for an assistant with limited memory.
search { query, max_results (1-10, default 6), when ("day" | "week" | "month" | "year") }
fetch { url, offset }What they do
searchreturns titles, addresses and short snippets, and tells the assistant to read the pages that matter, since a snippet alone is easy to misread.fetchreturns a page's main content as clean text, a part at a time, and says whichoffsetto pass to continue. Raw web pages are mostly markup, which would swamp the assistant, so the useful text is extracted. A PDF or Office file is saved and handed to the document reader.
Why they work well
| Design choice | Why |
|---|---|
| Every result is bounded and says how to continue | Results that are too long get cut without warning, so none comes close. |
| Addresses are treated as untrusted | The assistant picks addresses, and pages can be written by anyone. Nothing on the person's own computer or network is ever fetched. |
| Pages that need scripts are handled | Some pages show nothing until their scripts run, so these are opened in a hidden browser window. |
| Privacy is respected | The person is asked once per chat before anything leaves their computer. |
| Messy input is forgiven | Numbers sent as text, and "past week" or "7 days" for the time window, are understood. |
Using it from your skills
- Say when to search and what to keep: "search for each quote, fetch a quote page, and keep the price with its address and time".
- Ask for sources: "say 'as of' with the time and the source".
- Say what to do on failure: "never give a price from memory; if you can't find one, say so".