Are the classification numbers real or just demo numbers?

They are real measurements on a very small sample: four arXiv PDFs in tests/fixture with categories derived from arXiv subject tags, scored 4/4 on the Jev backend and 3/4 on laya. Nothing in the README suggests the rubric generalises to your document types, and there is no confusion matrix or per-category breakdown. Run with -n first and read the decisions yourself.

Will it reorganise my files destructively?

Not by default. The default is to create symlinks in the category folders, which leaves the originals exactly where they were; -m moves and --copy copies if you want that. -n prints the decisions without placing anything, so a first pass over a real folder costs one dry run. Name collisions are suffixed rather than overwritten.

Can it work without sending my documents to an API?

Yes, through the laya backend: pip install watfile[laya] for MLX on Apple Silicon, watfile[laya-coreml] for Core ML, or watfile[laya-torch] for any OS. No API key is needed, checkpoints download once and then run offline, and quoted speeds are roughly 13 ms per decision on MLX, about 5 ms on Core ML and 45 to 450 ms on CPU. Because laya's context window is only 512 to 1024 tokens, long extracts are split into chunks and re-classified until a decision is decisive, up to ten chunks.

How many calls and tokens does a folder of PDFs cost?

Batching is automatic on the Jev backend: about 256 tokens per document inside a 30k-token window, which the README puts at roughly 100 documents per call, so 100 files land near two calls rather than 100. --batch caps the batch size, --chunk-tokens sets the per-document budget and --no-batch forces one call per file for debugging. The laya backend does not batch at all.

How do I install watfile?

It needs Python 3.12+ and uv; run it one-off with uvx watfile, install it on your PATH with uv tool install watfile, and update with uv tool upgrade watfile or watfile --self-update.

How does watfile stop low-confidence files from being filed?

Placement is gated at a 0.5 confidence floor by default, and --min-confidence changes it with 0 disabling gating; raising the floor silently skips more files rather than flagging them.

CommunityJev Practices & Patterns

Watfile

Q: Which file types can watfile classify?

Text formats .txt, .md, .markdown, .rst, .log, .csv and .json are read directly; PDFs go through liteparse with only the first two pages parsed and OCR disabled, so scanned or image-only PDFs are skipped.

Python CLI that sends each document's text extract to a Jev Choice question and files the file into a matching category folder, with a local laya backend available for offline runs.

See highlights

Open github.com

Stars: 2
Language: Python
Published by: Community
Author: jexp
Last verified: Oct 2026

jexp/watfile

User repository on GitHub

View on GitHub

Text/PDF - File categorization and sorting with Typesafe AI Jev or local calibrated decision model

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Classify files with a decision-making AI model and sort them into category folders. watfile sends each document's text (title/abstract-grade extract) to a TypeSafe AI Jev (System One) Choice question, gets back a typed answer with a selected category, per-category probabilities and confidence, then moves the file into the matching folder. A Classifier abstraction keeps the backend pluggable — local MLX (laya) and other backends slot in later.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

The Jev backend packs documents into one system_one call at roughly 256 tokens each, up to about 100 files per 30k-token window.
Categories come from a comma-separated list with optional name:description hints, or from the subfolders of a target directory.
Dry-run prints decisions without touching anything; the default placement is a symlink, with -m to move and --copy to copy.
The laya backend runs locally on MLX, Core ML or PyTorch and uses adaptive multi-chunk classification because its context window is small.

Watch out

Accuracy rests on four arXiv fixture PDFs whose ground truth comes from their arXiv subject tags: 4/4 for Jev and 3/4 for laya. That is four documents, not a benchmark, and tests/fixture is where they live if you want to look. PDFs go through liteparse with only the first two pages parsed and OCR disabled, so scanned or image-only PDFs are skipped, and text inputs are limited to .txt, .md, .markdown, .rst, .log, .csv and .json. Files that cannot be classified are skipped with a warning and name collisions get a _1, _2 suffix, so a run can leave things behind quietly. Placement is gated at a 0.5 confidence floor by default, which means raising it silently skips more files rather than flagging them. Requires Python 3.12 and uv, and the key resolves from TYPESAFE_API_KEY, then .env in the current directory, then ~/.config/watfile/config.toml. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev_local

Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

Argos1111/jev_localPython

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.