Aअ

Indic Markdown Translator

Documents

Markdown, images of text, PDFs

Documents stay in this browser.

Untitled

From English → Into

Connect a model

Translation runs on a model server on this computer. Start it, then test the address. Nothing you paste leaves this machine.

Endpoint Model

The image’s language This page also has English

Source English Markdown

Source is typed in Latin letters Source looks like typed in Latin. Leave blocks already in Devanagari as they are

Only blocks to check Show Latin transliteration

Source English Translation Hindi

j k move · a accept · r retry · e edit

Result Hindi Markdown

Deliver

Add a note that a model made this

The review copy puts each source block beside its translation, for a reader who will check it.

Paste a document to begin.

Writing

No model and no server. Markdown in, Markdown out. This draft stays with the document until you use it as the source.

Input Latin Output Hindi

Settings

Model Endpoint

Any OpenAI-compatible server with /v1/completions: llama.cpp, Ollama, vLLM. Chat-only APIs, such as Sarvam's, do not work yet: the model is given a page of translations to continue, because asked in chat it answered the text instead of translating it.

Model name API key Remember the key on this device

The key is sent only to the endpoint above, and is cleared when you change its address. Without Remember, it lasts until this tab closes. Remembered, it is kept where your other pages at micahchoo.github.io could read it. Anyone using this browser can see it either way, so use a key with a spending limit.

Translation

Passage length Context blocks Retries

New documents start in

Each document keeps its own language. This is only the starting target for a new document.

Passage length is the most characters sent at once; a longer paragraph is split at sentence ends. Context blocks are earlier translations the model sees for consistency.

System prompt

{L} is replaced by the language name, {S} by the source label.

Examples

Pairs separated by a blank line. These teach the model the format; keep them general, not about one subject. They are saved for the language shown above only.

Front matter keys to leave untranslated End downloads with a note that a model made the translation Let the model check its own answers for a lost "not" or a related language. This makes translation about three times slower. llama.cpp servers only.

Storage on this device

Each downloads again when next needed. Documents and remembered words are not removed here.

Remembered words

When you edit a block, the page remembers your words for that source sentence in that language. The same sentence in a later document comes back with your words and is never sent to the model.