1992920 - History Semantic Search suggests unrelated items for Chinese

Status: RESOLVED FIXED

(Firefox :: Address Bar, defect, P2)

Description

[lilydjwg]

Attachment: 图片.png

In the screenshot the search term is "Japanese input method" in Chinese, and the results are "untitled document" and "cannot complete the request" in Chinese.

I have a lot of other cases but it contains private urls. Whenever there are not a lot of results, two sematic search results appear, but usually not semantically related, especially for Chinese.

Originally posted as bug 1991590 comment 1 but I'm advised to open a new bug for this case.

Comment 1

[Daniel Bodea [:danibodea]]

I am attempting to reproduce your issue on my system, but it would seem that I might need some help finding the cause.

These are the steps that I used to reproduce:

1. I launched the latest Firefox Release v143.0.4 AND Nightly v145.0a1 zh-CN language.
2. I set the Traditional Chinese language as the display language inside the browser.
3. Opened a new tab and input the same search term "日文输入法" in both URLbar and Searchbar.
   Notice: The suggestions displayed below appear to be related because the characters displayed are also displayed in the search term.

This being said, I cannot seem to reproduce it on my Windows 10 system using the steps above.

Please answer some questions to try and narrow down the cause:

1. Which operating system are you using?
2. In which Firefox version did you reproduce this issue?
3. Can you tell us which Add-ons you are using?
4. Does it reproduce in safe mode?
5. Does it reproduce in a newly created profile? (info here)

Thank you for your help! Please let me know if you need help with any of the instructions above.

Comment 2

[lilydjwg]

Notice: The suggestions displayed below appear to be related because the characters displayed are also displayed in the search term.

Good catch, I didn't realize that. But it's far from sematic anyway.

1. I'm using Arch Linux
2. I don't remember the exact version, but it might be 2025-09-28 judging from the date I took the screenshot
3. Well, I'm using about 60 addons
4. Yes. I disabled sematic search after this bug and it took quite some time to reindex my history. The version is 2025-10-07 this time.
5. No, sematic search seems to be not populated (there is no actual data in places_semantic.sqlite).

It seems that Firefox is using the all-MiniLM-L6-v2 model, which is small but doesn't seem to understand Chinese in a sematic way (and it is not listed in the Multilingual Models section). The BAAI/bge-m3 model could give relevant results but it's big. Maybe Firefox could use an external text embedding API (e.g. using llama.cpp) instead and let the user to choose and run a model. (Is a webextension already able to do that?)

Comment 3

[Marco Bonardo [:mak]]

For now the local model we use is English-based, it doesn't mishandle latin-based languages, but it's not ready for the rest of the world.
I'll implement a Region/Locale filter to limit the feature for now, and in the future we can extend it once we can get better local models (either multi-language, or specific based on the user history main languages).

Comment 4

[Marco Bonardo [:mak]]

Attachment: Bug 1992920 - Limit semantic history search to a list of supported locales. r=jteow

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/d0f27a2fe758
https://hg.mozilla.org/integration/autoland/rev/1422d02d0271
Limit semantic history search to a list of supported locales. r=jteow,urlbar-reviewers,cgopal

Comment 6

[Cristina Horotan [:chorotan]]

https://hg.mozilla.org/mozilla-central/rev/1422d02d0271