use remotely provided model name and owners when showing model choices
Categories
(Core :: Machine Learning: Frontend, enhancement)
Tracking
()
People
(Reporter: Mardak, Assigned: ngrato)
References
(Blocks 3 open bugs)
Details
(Whiteboard: [genai][other])
Attachments
(1 file, 1 obsolete file)
bug 2021947 current patch has a mapping of model choice ids (1, 2, 3) to model names and owners, e.g., 1 = gemini-flash-lite and google, so we can send it for telemetry. this is a refactoring of that user facing data added in bug 2021194 for firstrun
the names (without owner) are actually slightly different from what we landed for about:preferences bug 2006248 ("gpt-oss-120b" vs "gpt-oss-120B" and "Qwen3-235B-A22B-throughput" vs "Qwen3-235B-A22B")
and also different from the actual model name the server expects, e.g., qwen3-235b-a22b-instruct-2507-maas and gemini-2.5-flash-lite, which now is built in sync from remote settings bug 2022251. this remote setting dump could be used as a source of truth to address this bug but it's currently lacking the owner data
this bug is for updating the ui and telemetry where i believe even if the ui is wrong, since bug 2012580, we should correctly use a remotely changed model matching the user's modelChoice pref anyway
Updated•12 days ago
|
| Reporter | ||
Comment 1•12 days ago
•
|
||
for actual implementation assuming we can rely on remote settings records https://searchfox.org/firefox-main/source/services/settings/dumps/main/ai-window-prompts.json
we would probably need to select the records for feature == chat (?) and the corresponding model_choice_id and potentially filter out for a relevant version, e.g., "2.x" if the current version of firefox does not know about / support "3.x" versions. i see there's an is_default field and various others that may be relevant to check
of course edge cases of what to do with multiple matches, e.g., version 2.10 vs 2.11, or somehow record missing for a given model choice id
unclear what should happen if we start applying the user model choice pref to non-chat prompts… e.g., potentially we say fast/1 = chat system prompt uses gemini while title generation prompt uses <fancy-other-model> (instead of the current remote record indicating to always use qwen)
Updated•2 days ago
|
Updated•2 days ago
|
Description
•