ollama-batch-processor

Run whole books through a local LLM with Ollama: translate, prepare text for audiobook narration, strip book formatting, or paraphrase — chunk by chunk, file after file, with every step saved.

What it does

• Queue any number of .txt / .md / .srt / … files (or folders), press Start, get name_processed.txt next to each source (or in a folder of your choice).

• Pipeline of operations, run top to bottom — each with its own model and settings:

  Operation
  Translation	any language pair, meaning-first prompts with idiom handling; the end of the previous translation is fed back as context so names, terminology and register stay consistent across chunks
  Audiobook	spell out numbers, expand abbreviations, normalise punctuation for speech pacing, remove visual-only formatting
  Debookify	remove footnotes / page numbers / indexes, running headers & footers, normalise chapter headings
  Paraphrase	improve flow, simplify language, remove idioms, shift tone (formal / casual / professional / conversational)

  All checked tasks of one operation are merged into a single prompt, so the text passes through the model once per operation.

• Smart chunking: breaks at paragraph or sentence ends, configurable size / overlap or whole-file mode; the Ollama context window (num_ctx) is sized automatically from the chunk so nothing is silently truncated.

• Progressive saving: a .partial file grows chunk by chunk, each pipeline step is written to its own file, Stop keeps what is finished. Duplicate paragraphs at chunk seams are removed.

• Robust output: <think>…</think> blocks of reasoning models, "Here is the translation:" preambles, wrapping quotes / code fences are stripped; a warning is logged when a chunk's output length is suspicious.

• Server panel: URL, connection test, model list refresh, request timeout, keep-alive, num_ctx / num_predict / top_p.

• Presets (menu): Translate EN→CS / CS→EN, Audiobook preparation, Book cleanup, Plain language rewrite, Translate then audiobook — plus save / load / import / export your own and save as defaults.

• Prompts and operations live in config.json next to the app (File → Edit operations / prompts): add your own operation with options and a prompt and it shows up as a tab.

Install

Grab a prebuilt binary from the latest release (Windows / Linux / macOS, no Python needed), or run from source:

pip install -r requirements.txt
python main.py            # run.cmd / run.sh do the same

You need a running Ollama with at least one model:

ollama serve
ollama pull qwen2.5:14b   # or llama3.1, aya-expanse (great for translation), gemma3 ...

The server can be on another machine — put its URL in the Server tab.

Tips

• Translation: temperature 0.2–0.4, chunks of 2000–3000 chars, overlap 200. Bigger models translate noticeably better; aya-expanse and qwen2.5 are strong for European languages.
• Cleanup tasks (audiobook / debookify): temperature 0.0–0.2, larger chunks (3000–4000), overlap 0.
• Slow CPU models: set the request timeout to none in the Server tab.
• Out of memory: reduce the chunk size (this shrinks num_ctx) or set num_ctx manually.
• Output is <name>_processed.txt, steps are <name>_processed_step01_translated.txt etc.; existing files are skipped unless Overwrite is on.

Build

pip install -r requirements.txt pyinstaller && python build.py produces dist/ollama-batch-processor-<version>-<os>-<arch>. The GitHub workflow builds all three platforms on every tag and attaches them to the release; the Linux job installs Ollama, pulls qwen2.5:0.5b and runs a real translation through the frozen build (OLLAMA_BATCH_SELFTEST=<text file>).

Changes in 2.0.3

• The Windows taskbar shows the app's own icon. The taskbar button takes its icon from the process's Application User Model ID rather than from the window, and with none of its own the process was grouped under whatever launched it and wore that program's icon. One is now set before any window exists, and it carries no version number so a pinned button survives an upgrade.

Changes in 2.0.2

• No partial result files: outputs, step files and presets are written to a temp name and renamed into place only when complete; Stop or a server error removes the progress file
• Every process the app spawns is killed on quit or crash (never an Ollama server you started yourself)
• Adding files or folders rejects duplicates ("n already in queue")
• Natural sort order for folder scans and lists (chapter 2 before chapter 10)

Changes in 2.0

• New GUI in the whisperer style: queue with per-file status, live chunk progress and speed, log, presets menu, Server tab with connection test and model refresh
• Streaming requests in a worker thread (no more qasync / aiohttp); Stop really stops, mid-chunk
• Fixed: temperature was ignored for translation and passed raw (3.0 / 5.0!) to the other operations; config.json and the stylesheet were loaded from the current directory (broke when started from elsewhere); errors were written into the output as [ERROR: …] placeholders; the batch continued after Stop; output naming broke for non-.txt files; no num_ctx was set, so long chunks were truncated by Ollama's default 2k context
• Automatic context-window sizing and an automatic answer-token cap (a looping model can no longer run forever), keep-alive, timeout, <think> stripping, wrapper/prefix cleanup, length sanity check
• Translation continuation uses the previous translation (not the source) as context; chunk boundaries prefer paragraph breaks; duplicate removal only affects substantial repeats
• Input encoding detection (UTF-8 with BOM, UTF-16, cp1250/cp1252 fallback), overwrite guard, .partial progress file
• Standalone builds for Windows / Linux / macOS via PyInstaller + GitHub Actions