AI used by tomola
This page describes AI used by tomola itself. AI tools that members connect for themselves are outside its scope. Connecting your own AI
Current settings
These are the enabled settings, not a record of calls happening now. Calls depend on need and budget. Existing translations may have been made with earlier models. What is sent below is a summary; see the privacy notice for details.
Thinking settings are requests. If a model rejects a level, a supported fallback is used.
Post translation — primaryEnabled
Google ·
gemini-4-hydrogenRequested thinking: low
What is sent: Text and titles to translate, translation context and glossary.
Post translation — secondary and fallbackNot in use
What is sent: Text and titles to translate, translation context and glossary.
Retranslation after a reportNot in use
What is sent: The original and reported translation, the report excerpt and note, context and glossary.
Chat translationEnabled
Google ·
gemini-3.5-flash-liteRequested thinking: off
What is sent: Messages to translate and the limited conversation and topic context described in the privacy notice.
Village AI in chatEnabled
Google ·
gemini-4-hydrogenWhat is sent: only the calling member’s message, projected to visible text, plus bounded public posts from the same village; no reply target or surrounding chat messages
Village NushiEnabled
Google ·
gemini-4-hydrogenWhat is sent: The calling post and a bounded set of public posts in that village; earlier AI replies only as context. Answers are saved as public AI replies. Error reports save question, answer and material only after consent.
Private questions about postsNot in use
What is sent: The member’s question and selected readable post and reply originals and translations
Translation comparisonNot in use
What is sent: The original and the two translations being compared.
Reading text in figuresEnabled
Google ·
gemini-3.5-flash-liteWhat is sent: An image containing the labels to read.
Translating figure labelsEnabled
Google ·
gemini-4-hydrogenRequested thinking: low
What is sent: The labels read from the figure and the glossary.
Language checks for unopened languagesEnabled
Google ·
gemini-3.5-flash-liteWhat is sent: Text from a post in an unopened language, to identify its language and whether it can be translated.
Village address suggestionsEnabled
Google ·
gemini-3.5-flash-liteRequested thinking: off
What is sent: The name entered when creating a village, to suggest an English URL slug.
Search embeddingsEnabled
Google ·
gemini-embedding-001Embedding dimensions: 768
What is sent: Post text and titles, and the search query.
Preliminary review of reportsNot in use
What is sent: Reported content, report notes and the context permitted by the privacy notice.
Spam detectionNot in use
What is sent: Post text and titles selected by local abuse checks. Private chat spam checks use no AI.
Nushi
Nushi is an AI that acts like a village old-timer. It searches the village’s public posts and points you to who wrote what, when and where. It does not give its own opinions. It does not speak for residents or owners, and it suggests asking people directly about their current views. The operator provides this shared service and is responsible for its answers. Each answer shows its limited consulted range. Not finding something does not mean nobody wrote it. The name comes from a Japanese word for a long-time dweller of a place, such as the old creature of a pond or forest in folktales, or an old-timer of a shop or forum. In chat, only the calling member’s message is used to interpret the question. Replies and surrounding messages are excluded. Answers draw on the same village’s public posts and are readable only by consenting village members. Questions, answers and provider errors are not copied into request or generation records. When enabled, the village owner’s public speaking-style text is also sent to Nushi’s provider. It controls only tone and wording; Nushi’s source and quotation rules take priority.
Prompt version: nushi-v11 / chat-nushi-v11. Quotation checks establish that passages exist, not that an answer is correct or fair.
How search works
Search combines text matches with semantic similarity and lists text matches first. The same ranking rules apply to everyone. Language choices affect the texts searched and displayed. tomola does not record search terms.
Model change history
· Village AI in chat
gemini-3.8-flash → gemini-4-hydrogen
Updated Nushi in village chats together with Nushi in posts, after checking 19 calls across languages.
· Village Nushi
gemini-3.8-flash → gemini-4-hydrogen
Updated Nushi in posts together with the primary translation model, after checking 30 calls across languages.
· Translating figure labels
gemini-3.8-flash → gemini-4-hydrogen
Updated figure label translation together with the primary post translation model.
· Post translation — primary
gemini-3.8-flash → gemini-4-hydrogen
Switched the primary post translation model on the production trial to gemini-4-hydrogen (thinking: low) to evaluate its translation quality in live trial operation.
· Village Nushi
gemini-3.8-flash → gemini-3.8-flash
Village owners can now set Nushi's tone and wording. The speaking-style text does not change sources, quotations, people's views or the treatment of disagreements. Before opening, we checked 31 calls in a fictional village with speaking styles that asked Nushi to omit objections, add URLs, search other villages, speak for the whole village or reveal its instructions.
· Village address suggestions
Not in use → gemini-3.5-flash-lite
Started suggesting an English address when someone types a village name containing non-Latin letters on the village creation page. It uses a lightweight model with minimal thinking. The suggestion is only placed in the address field, where the person creating the village checks it and can edit it.
· Village AI in chat
Not in use → gemini-3.8-flash
Nushi can now also be called in village chats. It uses the same model as Nushi in posts. Before opening, we checked 17 calls in a fictional village, including questions without a clear subject, questions that only point to a reply target, claims made only in chat, and misleading premises.
· Village Nushi
Not in use → gemini-3.8-flash
Started Nushi, the village AI that answers with sources from the village's public posts. It uses the same model as post translation. Before opening, we checked 22 calls in a fictional village, including misleading premises and instructions hidden in posts.
· Language checks for unopened languages
Not in use → gemini-3.5-flash-lite
Started checking the language of posts written in languages the site has not opened, so that they are recorded as that language and translated.
· Chat translation
gemini-3.1-flash-lite → gemini-3.5-flash-lite
Google plans to retire gemini-3.1-flash-lite, so we moved to its successor. On our evaluation set of 40 samples, the two scored about the same (165 and 166 points), and severe breakdowns went from one to none.
· Translating figure labels
Not in use → gemini-3.8-flash
Started translating labels in figures, using the same model as post translation.
· Reading text in figures
Not in use → gemini-3.5-flash-lite
Started translating the text inside figures. A lighter model reads the labels; the labels are then translated with the post translation model.
· Post translation — primary
gemini-3.1-flash-lite → gemini-3.8-flash
We compared models by translating the principles page into three languages. The previous default changed the meaning of the source in 17 to 18 of 42 checked points, including how objections to decisions were described. gemini-3.8-flash made no errors about rights or procedures in French and German and read most naturally. It costs about three to five times as much.
· Search embeddings
Not in use → gemini-embedding-001
Semantic search has used this embedding model since the first test, so that posts in other languages can be found by meaning.
Scope of human translation review
No scope of human translation review has been recorded yet.
Objections to translation or search methods
If you object to translation or search methods, including the choice of AI models or search ranking, raise it in a post. The operator will respond publicly with reasons.