[Models] [Search Handoff] Assistant is making 15+ search toolcalls in a row with select queries esp w/ duckduckgo
Categories
(Core :: Machine Learning: General, enhancement)
Tracking
()
People
(Reporter: tburrell, Assigned: again)
References
(Blocks 1 open bug)
Details
(Whiteboard: [genai][models])
Attachments
(1 file, 3 obsolete files)
Assistant makes 15+ search toolcalls in a row for this query: https://mozilla.slack.com/archives/C0ADUFX0QRW/p1771523842601919?thread_ts=1771520791.301559&cid=C0ADUFX0QRW
This might be a problem with other search providers as well.
Updated•6 months ago
|
Comment 1•6 months ago
|
||
potentially content from duckduckgo with very long links is confusing the llm? this is currently doing pageExtractor.getReaderModeContent bug 1990614
| Reporter | ||
Comment 2•6 months ago
|
||
this is easy to reproduce by asking to search for multiple things, e.g., “find me shoes and seattle weather”
Comment 3•6 months ago
|
||
Check function description for get_open_tabs:
"Access the user's browser and return a list of most recently browsed tabs. Each tab is represented by a JSON with the page's url, title and description if available. Default to return maximum 15 tabs."
May be related.
Updated•5 months ago
|
Updated•5 months ago
|
Guard duplicate run_search tool calls within the same user message turn in fetchWithHistory(). When the model requests run_search after one has already executed for the current turn, the guard returns an error tool result instead of executing the search and navigating the browser. The model continues generating text and can search again on the next user message.
Updated•5 months ago
|
Updated•5 months ago
|
Updated•5 months ago
|
Comment 10•5 months ago
|
||
| bugherder | ||
Comment 11•5 months ago
|
||
Verified as fixed in our latest Beta 150.0b1 and our latest Nightly build 151.0a1 (2026-03-26), I tried this with Bing, Google as well as duck duck go and I'm getting suggestions each time without any search hand-offs.
Description
•