Humanitext Antiqua

Guide

1. Service status

The core Service is free for everyone: within the daily allowance there is nothing to pay. Answers are generated through hosted models and grounded in the indexed corpus.

To keep the service free — covering model costs and applying a fair daily allowance per person — the chat asks you to sign in with a Google account. What is recorded and how it is handled is described in the Privacy Policy.

The allowance is a daily amount of processing, resetting at 00:00 UTC (09:00 Japan time); the remaining share is shown at the bottom of the chat. How many questions that comes to depends on your settings — more with a lighter model and fewer passages, fewer with a premium model or many passages. Because the settings change how fast the allowance is consumed, you can spend it where it matters to you. There is also a service-wide daily processing ceiling: on a day when it is reached, no further questions can be accepted until the next reset. Both keep the running costs of a research project bounded, and both may be adjusted as operating conditions change.

2. Genre, author, and work filters

  • Genre (major / minor): restrict retrieval by literary genre. Choose one or more major categories to narrow the minor options; either may be left empty. Selecting a genre also narrows the author and work lists.
  • Authors: limit retrieval to one or more authors. When empty, the whole corpus is searched.
  • Works: once authors are chosen, narrow further to specific works.
  • Exclude the selection: invert the author/work selection to search everything except those entries. The genre filter always restricts, independently of this toggle.

3. Model and number of passages

There are three models to choose from, and all of them are available on the free plan. The difference shows up as arithmetic rather than access: a more capable model costs more processing per question, so the same allowance covers fewer of them.

  • Economical (gemini-3.5-flash-lite, the default): light and fast, and enough for most questions.
  • Balanced (gemini-3-flash-preview): it reasons further, which suits harder questions and longer passages of writing. It costs roughly 1.5x as much processing per question, so it uses up your daily allowance correspondingly faster.
  • Premium (gemini-3.8-flash): it reasons the most deeply, which suits intricate comparisons and long, careful prose. It normally costs roughly 4x the processing of the Economical model; through the end of 2026 the provider’s launch discount halves that, so the free allowance covers about twelve questions a day (about six from 2027).

The number of passages controls how much context is retrieved per question: 1–5 (default 3) on the free plan, 1–10 (default 5) on the paid one. More passages improve coverage. They also stay in the conversation: a passage retrieved for one question is carried forward as context for the rest of that conversation, so a higher number uses up your daily allowance proportionally faster.

4. Output modes

  • Q&A: concise answers with precise references.
  • Study: comprehensive, accessible explanations with background and translated evidence.
  • Research: in-depth philological analysis and interpretive commentary.
  • Conversation: speak with a chosen ancient author or character.
  • Custom: supply your own system prompt for specialised tasks.
  • No-context mode: answer from the model's own training knowledge, without retrieving texts — useful for translation, brainstorming, or broader context beyond the corpus.

5. The inquiry screen

  • Type a question and press Enter (Shift+Enter for a new line).
  • Answers stream in with numbered citations — click a number to jump to the passage it rests on. The Sources panel shows each passage in full: original text, English rendering, edition, and a citation you can copy. Open the reading view for a facing-page spread of original and translation.
  • After each answer, follow-up suggestions propose where the inquiry might go next; use Regenerate to re-answer the same question.
  • Inquiries are kept in your browser's local history — switch between them, delete them, or start a new one at any time. What is kept on the server is set out in section 6.
  • Use Export to download a conversation as Markdown or JSON, citations included.

6. Conversation history, and what the server keeps

The inquiry history shown on screen is saved inside your browser(localStorage). It is therefore specific to this browser on this device: it survives reloads and restarts, but it does not sync across devices, and clearing the browser's site data (or ending a private-browsing session) deletes it.

Separately from that, two kinds of record are kept on the server.

  • Usage statistics (every request): the corpus scope selected, the mode, processing time, approximate token counts, success or failure, and the language setting. No part of your questions or the answers, no account identifier, and no IP address.
  • Conversation records (only if you consent to research use): your questions, the generated answers, and the sources cited. This is off by default and can be switched either way in the settings at any time. Leave it off and no text is ever stored, with no feature withheld from you. If you turn it on and later stop, you are asked at that moment whether to keep the records already stored or delete all of them; records you keep can still be deleted from the account screen at any time.

Clearing your browser history does not delete the server-side records. For what they contain, how long they are kept, and how to have them deleted, see the Privacy Policy.

  • The most recent 30 inquiries are kept; beyond that, the least recently used are removed automatically.
  • To keep an inquiry permanently, use Export (Markdown or JSON) — citations included.
  • Within one conversation, all earlier turns and the passages retrieved for them are sent to the model as context. Nothing is silently dropped; once a conversation passes a certain length, you are invited to start a new inquiry instead.

Add it to your home screen

On a phone or tablet you can add this site to the home screen and launch it full-screen, without the address bar. Chrome on Android offers it as “Install”. iPhone and iPad show no prompt, so open the Safari share menu and choose “Add to Home Screen”. It stays an ordinary web page either way, and it does not work offline.

Feedback

Suggestions and issue reports are welcome via this form.

Acknowledgements

  • Hagiwara Foundation for Academic Research, 2023 (4th Research Grant)
  • Mitsubishi Foundation, Humanities Research Grant, 2024
  • Joint Support Center for Data Science Research, ROIS-DS-JOINT FY2024-1

Data sources

The corpus draws on the Perseus Digital Library and the Open Greek and Latin project (First1KGreek), together with works edited by the Humanitext Planetariuminitiative. Texts are re-segmented and normalised inside the Humanitext pipeline; author/work names, genre classification, and citation patterns are curated by the Humanitext team. The complete list of indexed works is available in the Corpus tab.

Credits

  • Naoya Iwata (Nagoya University / National Institute of Informatics)
  • Ikko Tanaka (J. F. Oberlin University)
  • Jun Ogawa (University of Tokyo)

References

  • Naoya Iwata, Ikko Tanaka, Jun Ogawa, “Utilizing Large Language Models for Research and Education in Western Classics” (in Japanese), Proceedings of the Annual Conference of JSAI, 38, May 2024. DOI: 10.11517/pjsai.JSAI2024.0_1N4OS1805
  • Naoya Iwata, Ikko Tanaka, Jun Ogawa, “Improving Semantic Search Accuracy of Classical Texts through Context-Oriented Translation”, IPSJ SIG Computers and the Humanities Symposium, 2024.