Introducing DocJev - a lightning-fast OSS library for document classification and splitting with jev ⚡️ Give a document alongside some natural language category rules. Jev will predict the document category (classify) or the boundaries between sub-documents (split). It is 6x faster than gpt-5.6-luna with equivalent accuracy. You can choose... between different OCR backends: liteparse: the fastest/most accurate free+OSS text parser out there. Perfect for digitalized documents. The speedups vs. luna above are inclusive of liteparse times. LlamaParse: the most advanced VLM-based solution for complex docs. Using this will of course add latency on document preprocessing. But it’s a good option if you want to cache the digitalized representations of documents for a variety of downstream tasks. Check it out: If you want to dive deeper into liteparse or LlamaParse check the links below: liteparse: LlamaParse:show more

Latency benchmarks:

You can find the full interactive report here:

Splitting docs that fast is useful

Nice. Would be very cool to have model routing based on PDF contents, e.g. use Marker if there are tables, etc.

need to try this 👀

The "natural language category rules" is the part that will get interesting fast. Looking forward to trying it out.

Natural-language rules as the interface are the right abstraction - humans own the taxonomy, Jev owns the cheap decision.

Nice example of keeping the interface simple while exposing both classification and boundary detection. The split output should make it easier to audit where a document needs review, especially when the classifier is uncertain.

Natural language category rules removes the whole annotation/fine-tuning step for document classification.

Just wired LlamaParse into a procurement RAG ingest last week. Tables + scanned PDFs were the whole problem. Caching the parsed doc and then classifying/splitting fast is exactly the missing step. Trying DocJev on the next batch.

Could you benchmark the scanned PDFs in the synthetic set next? The report's ~6x figures exclude OCR. Timing the full pipeline would show how much of that speedup reaches someone uploading a scan.

a bad split never throws an error. it just quietly changes every answer downstream and the citation still looks fine.

This is great

DocJev 6x on doc split. that's the hit

The split is the decision that matters, the classification is just filing. I've watched document pipelines where a perfect classifier sat on top of bad splits and the RAG answers were still garbage, because one wrong boundary means two contracts live in the same chunk. Get the boundaries right and the labels almost don't matter.

Important qualifier: the repo's accuracy pilot is a 40-PDF, 8-packet convenience sample with shared source/template families and uncontrolled provider caching. The reproducible report does not establish general split accuracy. Keep boundary review before unattended use.

Document classification and splitting, nice launch

ocr tradeoffs are the whole game — latency vs accuracy is what makes doc pipelines ship or stall.

文档分类/切分做成开源原语，比再堆一个「全能 RAG」更落地。DocJev 这种薄库才容易嵌进现有流水线。

Will the LlamaParse preprocessing latency bottleneck break real-time API limits if downstream tasks never actually reuse those cached representations?

the subdocument boundary bit is super useful, thats usually where doc classifiers get weird

fast splitting + classification is what unlocks unmonitored pipelines. tested an autonomous pipeline pulling macro data for brazil from outside: using fast models for schema boundaries instead of giant context windows cut latency 85% and completely eliminated 3am timeouts.

splitting docs from english category rules, no classifier zoo

The split between classification and boundary detection is a nice touch. OCR pipelines often get treated as one problem, but the backend choice matters once documents stop being clean scans.

huge！

how does DocJev handle edge cases where documents don't fit typical rules?

Similar picture on papers instead of documents: 1,000 arXiv abstracts, 24 categories, 57 ms each. With Opus 5 as judge Jev matched 85 of 100 at 153x lower cost per paper. Confidence under 0.90 flagged 14 of the 15 misses.

Turning the annotation step into a rules interface is a very good trade for this kind of workflow.

Is jev on your leaderboard? Curious on your benchmark accuracy vs latency for Jev vs others

This is a strong abstraction. Letting the user describe boundaries in natural language removes a lot of brittle preprocessing, especially when document formats keep changing.

docjev for classify + split with natural language rules is neat 6x faster is a serious flex on doc pipelines

I tested DocJEV with both a local model and the official JEV model — works equally great! 🔥