github felladrin/MiniSearch 2026.11.19.1

54 minutes ago

2026-11-19

  • When no voice is chosen in Voice Settings, the local engine now picks the best-sounding voice of the answer's language instead of the first one in the catalogue: medium quality first, then low, then x-low. Portuguese, for example, now gets pt_BR-faber-medium where it used to get pt_BR-edresson-low, at the same 63 MB download. High-quality voices are picked only when a language publishes nothing else, since they cost up to 137 MB against the 63 MB of a medium one. And when the answer is in the browser's language, a voice of the browser's region now wins among the ones of the same quality, so an en-US browser gets an American voice.
  • Raon-1 nano 3B (2083 MB, Apache-2.0, Korean and English) joins the Wllama catalog, labeled "(wllama support pending)": its architecture is not yet in the llama.cpp build the browser engine ships, so it cannot answer until upstream support lands (ggml-org/llama.cpp#32200 is open). The sharded copy is published and the entry is ready, so it starts working the day the engine updates.
  • A follow-up answer now reads the pages behind its own new results, as the first answer does, instead of answering from their titles and snippets alone. A page both searches returned also gets an excerpt picked for the follow-up question, where it used to keep the one picked for the first question.
  • The search query of a follow-up question now names what the question refers to: when the previous answer says Marvel moved Avengers: Doomsday to December 2026, "why did they delay the release" now searches for "Avengers: Doomsday release delay reasons" instead of a query that names nothing. The query is also written in the language of the question more often, because the model no longer sees the search results while writing it. Measured with Gemma 3 4B in the browser on seven conversations, three runs each: the query kept the subject in 21 of 21 runs (16 before) and the question's language in 20 of 21 (16 before).
  • The six search results an AI answer reads are now the six with the best relevance score, for the first question and for follow-ups. The result list still shows the search engine's first result at the top, but that result no longer takes one of the six places when its score is low.
  • A follow-up question with a relative date, such as "what comes out next month", now searches with the month and year written out ("December 2026"), as the first question already did, so the results match the right period.
  • When a search was still waiting for permission to download the in-browser model and a second one started (for example, by accepting the "Enable AI Responses?" prompt with AI responses already on), allowing the download loaded the model twice at the same time, and both answers failed with a file-lock error. Now only the newest search loads the model.
  • A follow-up question with a relative date ("who won it last year", "明年的呢") now keeps that date in its search query more often, so the date anchor turns it into the right year instead of the model writing a year itself ("2023" for last year). Measured with Gemma 3 4B in the browser on three such questions, three runs each: the relative date reached the anchor in 6 of 9 runs, against 3 of 9 before.
  • The search query of a follow-up question no longer keeps the quote marks some models wrap it in (“ ”, 「 」, « »), as it already did with straight quotes.
  • The search query of a follow-up question in Chinese, Japanese, Korean, Thai, Greek or Hebrew now stays in that language: when the model writes it in another script, the query is translated back, and when that fails too, the typed question is searched. Measured with Gemma 3 4B in the browser, "它大概多少钱" after a question about Apple's new phone searched in Chinese, naming the phone, in 9 of 9 runs; before, the query came out in Turkish, Persian or English in most runs. A question in another script (Latin, Arabic, Cyrillic...) whose query changes script also searches for the typed question now.
  • A follow-up question with a relative date now searches for the date the user meant even when the model writes a year of its own: "明年的呢" after a question about Apple's new phone used to search "苹果 2025 新手机 发布时间", and now searches "苹果 新手机 发布时间 2027年". The year comes from the question as typed, and a year the user wrote is kept. Measured with Gemma 3 4B in the browser on three such questions, nine runs each: the right date reached the search in 27 of 27 runs, against 17 of 27 with the old code on the same queries.

Don't miss a new MiniSearch release

NewReleases is sending notifications on new releases.