Skip to main content

Attaching an EPUB to an AI chat box only translates a few pages: why full books get cut off and how to translate a whole novel with its table of contents intact

Tri Thu知秋

On this page

ChatGPT can read the text inside an EPUB file to summarize the plot, look up details, or test-translate a few short passages, but it cannot translate an entire EPUB novel of hundreds of chapters and repackage it with the table of contents intact in one chat turn due to per-reply output length limits and the multi-file compressed structure of EPUB. Translating a full book while keeping chapters, illustrations, and navigation links requires a batch workflow with EPUB write-back.

In a sunlit guofeng mechanical hall bathed in warm golden afternoon light, a young artificer in celadon jade and warm cream robes stands before a wide stone table: on the left, loose paper sheets scatter in the wind from an overloaded small furnace, while on the right, guided by cinnabar-red silk threads under his hand, a brown-bound ancient tome lies flat with bronze gears and a celadon jade frame holding every page, illustration, and table-of-contents tab intact.
A general chat box works well for quick Q&A and short excerpts, while translating a full EPUB novel requires unpacking the book spine and writing translated text nodes back into the original package.
Your goalAttaching an EPUB in the ChatGPT chat boxBatch EPUB translation on NovelVerba
Summarize a chapter, look up plot details, or test-translate 1 to 2 pagesPractical and fast: extracts text from the file to answer questions right in the conversationNo need to create a project if you only want a quick preview before reading
Translate a full novel of hundreds of chapters (up to 5,000,000 source characters)Not feasible in one turn: replies stop after a few pages or collapse the story into a summaryRuns in batches of up to 50 chapters or 200,000 source characters, saving each finished chapter immediately
Keep the table of contents (toc.ncx / nav.xhtml), reading order (spine), and illustrationsLoses package structure: outputs plain text in the chat window or a stripped text fileWrites translations back (write-back) into the exact text nodes inside the original ZIP package, keeping image bytes and inline tags
Keep character names and forms of address consistent across the whole bookNames and pronouns drift easily when split across dozens of separate chat sessionsLocks a Names & terms dictionary (up to 500 entries) at the start of every batch
Verify no paragraphs were dropped before sending the book to an e-readerNo automated line-count check between source chapters and translated outputChecks source match and complete translated line count per chapter before exporting monolingual or bilingual EPUB

What actually happens when you attach an EPUB file to the ChatGPT chat box?

Almost every web novel reader who tries AI translation starts with the same experiment: you take a Chinese, Japanese, or Korean .epub file, drag it into the ChatGPT chat box, and ask it to translate the whole book into English and give back a downloadable EPUB. Seconds later, instead of a complete translated novel, you get a plot summary of the opening chapters, or two to three translated pages followed by a prompt asking if you want to continue. Even when the chat window generates a downloadable file, opening it reveals a short text fragment stripped of cover art, illustrations, and chapter links.

To be fair to general chat tools, this happens not because ChatGPT cannot read your file. If your goal is simply to summarize a chapter, check when a side character first appeared, or test-translate one or two pages to sample the writing style, attaching the file or pasting a passage into the chat box is convenient and works well. A chat interface is built to extract text from uploaded files for quick Q&A.

The wall appears when you ask a chat window to turn a book-length file into a complete translated book. Every conversational AI interface operates under an output length ceiling: while it can read in a large file, a single reply can only write out a few thousand words at a time. A typical web novel runs from several hundred thousand up to 5,000,000 source characters across hundreds of chapters. Forced to answer within one chat turn, the AI must either condense descriptions into a summary or halt after a few pages.

Why can't an EPUB file be translated and stitched back together like plain text?

Even setting aside single-reply output limits, an .epub file is not a flat stream of text like a .txt document. Under the hood, an EPUB is a structured ZIP archive containing META-INF/container.xml, which points to an .opf package file declaring every resource in the book (manifest) and the exact linear reading order of chapters (spine), alongside navigation files (toc.ncx or nav.xhtml), hundreds of .xhtml chapter files, CSS stylesheets, and images.

If you unzip an EPUB by hand, copy chapter text into a chat window, and paste translations back in, a single broken HTML tag or altered path can break the table of contents, strip bold and italic formatting, or leave blank pages on an e-reader. The table below compares what happens to each part of the EPUB package when copying through a chat window versus using the write-back mechanism on NovelVerba.

EPUB package componentManual copy-paste or chat-box exportEPUB write-back on NovelVerba
Package routing (container.xml, .opf, spine)Ignores the spine order or outputs a flat document that breaks internal chapter pathsReads container.xml and the OPF file to split chapters in the exact reading order of the original spine
Table of contents (toc.ncx / nav.xhtml)Chapter links break when text is copied out and reassembled by handPreserves internal file targets inside the original ZIP archive so TOC links land on the right chapters
Chapter files (.xhtml) and inline tagsOften drops <b>, <i>, <a>, or <br> tags, or produces malformed XMLReplaces only text inside text nodes while preserving inline tags and paragraph whitespace
Illustrations and non-text assetsCover art and chapter illustrations are lost when extracted as chat textCopies every image and non-text asset byte for byte from your uploaded EPUB archive
Pre-export completeness checkpointNo check to catch paragraphs silently skipped in the middle of a chapterVerifies that source content matches the translated version and that the translated line count is complete

Instead of building a new file from loose text, NovelVerba's write-back mechanism reopens the exact ZIP archive you uploaded, locates each paragraph by the positions recorded at split time, and swaps only the text-node contents with the translation. Before export, every chapter passes a checkpoint verifying that source content matches the translated version and the translated line count is complete; if a chapter is incomplete or mismatched, the system reports an error and refuses to export a broken file. Full details are in How to translate an EPUB novel into English and export a file you can read.

If you have an external .epub file and want to check whether its ZIP archive, spine, table of contents, or XHTML tags have structural errors before reading or translating, drop it into EPUB Validator: Check EPUB Files Before Reading or Translating. It checks 12 structural points of the EPUB package right in your browser without uploading the file to a server (not equivalent to EPUBCheck).

What happens to character names and address terms if you paste chapters one by one into a chat box?

To bypass the single-reply length limit, some readers copy out one chapter at a time and paste chapters into a chat box across dozens of turns. Across a 300-chapter novel, that manual loop hits a third wall: context drift between chat sessions. Without a locked glossary injected into every run, a character name or cultivation realm transliterated in chapter 4 may be translated literally or spelled differently when you open a new chat window at chapter 19.

Even more jarring than shifting names is shifting forms of address and hierarchy. Chinese, Japanese, and Korean novels frequently omit pronouns or rely on kinship and sect titles. When chapters are translated in isolated chat turns without pre-locked relationship rules, two martial brothers or a master and disciple can shift tone or flip seniority mid-dialogue. How to lock relationship-based address terms before translating is covered in Choose address terms by character relationship when translating Chinese novels.

Rather than hoping a chat window remembers 300 chapters on its own, NovelVerba stores that memory in the Names & terms table (up to 500 entries), locking one dictionary set at the start of every batch and saving each finished chapter immediately to the database. How to pick up at the next untranslated chapter and review the name table between batches is explained in Translating a long novel in batches: where the next session picks up, and whether the names still match.

What is the 5-step workflow to translate a full EPUB novel and keep its table of contents?

Turning a raw .epub novel into a finished translation with working chapter navigation and intact illustrations follows five steps. A single upload accepts an EPUB file up to 15 MB and up to 5,000,000 source characters per novel, while translation runs in batches with upfront source-character pricing.

  1. 1. Upload your EPUB file (up to 15 MB and 5,000,000 source characters): the system opens the ZIP archive, reads container.xml and the .opf file, and splits chapters in the exact order of the book's spine.
  2. 2. Scan and lock the Names & terms table on the first 5 to 10 chapters: the AI extracts character names, places, titles, and terms so you can review and lock up to 500 entries before translating.
  3. 3. Select your first translation batch (up to 50 chapters or 200,000 source characters): approve the upfront credit quote calculated by source characters; each finished chapter is saved immediately to the database.
  4. 4. Run subsequent batches with Continue translating and review flagged lines: check new characters in the name table at each batch boundary, and fix flagged lines inline or retranslate specific chapters.
  5. 5. Export the translated EPUB or bilingual EPUB: the system verifies source match and complete line count for every chapter, then writes the translation back into your original EPUB package.

How do you send the exported EPUB to a Kindle or Kobo e-reader?

When exporting via write-back, NovelVerba updates the dc:language metadata tag to match your project's target language so your e-reader shelves the book under the right language. If you want to compare original and translated text side by side, choose Bilingual EPUB: each source paragraph sits right above its translation with a gray left border and smaller font attached directly to the paragraph, rendering cleanly even on e-readers that ignore external CSS.

Three layout specifics of Chinese and Japanese books are deliberately relaxed during write-back export so horizontal text reads naturally: vertical text (writing-mode: vertical-rl) is removed; furigana wrappers (ruby) are unwrapped while keeping the translated base text; and decorative initial letters (drop-cap) keep their wrapper only when the translation is a single character so a full word is not blown up. Older EPUBs encoded in UTF-16 are refused at write-back export to avoid character corruption (convert to UTF-8 first or use TXT export).

Once you download your translated EPUB, see Reading translated novels on an e-reader: choosing the right EPUB export, checking errors, and fitting your screen for choosing between original-layout and newly built EPUB exports, fixing blank pages, and fitting margins to Kobo, Boox, Xteink, or Kindle screens. To transfer your translated EPUB to a Kindle via Send to Kindle without delivery errors, see How to send a translated EPUB novel to Kindle and check for file errors before sending.

Upload your EPUB file and translate the first batch