LlamaIndex compares Jev, Qwen and open encoders on document-processing decisions. The linked notebook uses LiteParse to supply text and page signals before models choose labels, page boundaries or a parsing tier.
Original by LlamaIndexEvaluationIntermediate16 min 48 secPublished
Separate extracting text from deciding what to do with it. LiteParse supplies text and page signals; the models make choices about that material.
Compare models on the decision you need: language, orientation, document type, page boundaries or whether to use a heavier parser.
Include preprocessing in the comparison. The orientation example checks text from four OCR rotations; parser triage assumes orientation has already been fixed.
Worth knowing
First-party comparison, not an independently repeated benchmark. Results and costs depend on the chosen samples, warmed model containers and rough cost attribution. Four-rotation OCR adds work outside the decision call. The notebook labels parser upgrades using OCR character error against known source text; it does not prove which paid parser will recover a page. Jev receives text and signals, not the original document images.