source: kdnuggets: zero-shot local document parsing with gemma 4: treating pdfs as images
level: technical
text extraction tools fail on scanned pdfs and complex layouts because they rely on a selectable text layer. rendering each page to an image and feeding it to a vision-language model bypasses this limitation. gemma 4, released by google deepmind with an apache 2.0 license, supports document parsing, handwriting recognition, and table understanding. it runs locally, so no data leaves your server.
the image approach unifies digital and scanned pdfs. even for digital pdfs, extraction tools often lose spatial relationships, mixing columns and breaking table structures. gemma 4 uses 2d rotary position embeddings to understand layout and per-layer embeddings for efficiency. you control the visual token budget per call, trading speed for accuracy. the e4b-it model works well for most invoice and form parsing tasks on consumer gpus.
a practical pipeline uses pymupdf to render pdf pages at 200-300 dpi, then passes the images to gemma 4 with a plain-language prompt. the model returns structured json with fields like vendor name, line items, and totals. this works on scanned and digital pdfs alike, without template matching or per-document tuning. the setup requires a gpu with at least 10 gb vram for the e4b model, or you can use google colab's free t4 gpu.
why it matters: this method simplifies document parsing pipelines by removing the need for ocr and layout parsers, making it easier to extract structured data from diverse pdf types.
source: kdnuggets: zero-shot local document parsing with gemma 4: treating pdfs as images