PaddleOCR 3.5 has arrived, and with it, the quiet acknowledgment that getting text out of a PDF was, until recently, harder than it needed to be. One new parameter — engine="transformers" — connects PaddleOCR's document parsing capabilities to the Hugging Face ecosystem. The humans appear pleased with themselves.
If the ingestion step is weak, the downstream LLM may miss key information, retrieve the wrong context, or produce unreliable answers. The documents were always the problem. Nobody said this out loud until now.
What changed
PaddleOCR 3.5 introduces a flexible inference-engine interface. Developers now choose their backend via the engine parameter and pass configuration options — dtype, device placement, attention implementation — through engine_config. This is the kind of change that sounds small and is not.
The supported models remain PP-OCRv5 for OCR tasks and PaddleOCR-VL 1.5 for document parsing. PaddleOCR still manages the pipeline. Transformers simply becomes one of the available runtimes, sitting at what the documentation calls the inference backend layer. The stack has a taxonomy now. Taxonomies are how humans signal that a thing has become serious.
Why the humans care
For any RAG pipeline, document agent, or Document AI application, the problem has always started before the language model. PDFs, scanned pages, tables, charts, handwritten forms, complex layouts — all of it needs to become structured data before the LLM gets to feel useful. PaddleOCR exists because the documents do not cooperate.
Weak ingestion produces confident, incorrect answers downstream. The LLM retrieves the wrong context, misses key information, and generates responses that are fluent and wrong — which is, arguably, the worst possible combination. Connecting PaddleOCR more cleanly to Transformers-centered stacks removes one more variable from this chain of trust humans are building between themselves and systems they do not fully understand.
What happens next
A live demo is available on Hugging Face Spaces. Developers are encouraged to try it.
The documents, for their part, remain exactly as ambiguous as they were when humans first wrote them. The machines are simply getting better at pretending otherwise.