My system takes different PDF files uploaded by users every day (each file is unique and used only once). For each PDF, Gemini should generate a structured JSON array of objects with fields like title, trade, page numbers, image indexes, etc.

This works flawlessly for normal reports (~80–120 pages).
But when the PDF is large (200–250+ pages) — especially image-heavy — Gemini breaks:

• Returns empty or partial JSON
• Ignores most pages
• Produces invalid schema output
• Sometimes seems to “not see” the document at all

Important details:

• PDFs are often small in file size (e.g., 100–300 KB), but page count is huge (goes from 1 to 400 pages pdf files)
• The files contain many images/photos
• I pass the PDF via inlineData or JSONL for batch prediction
• Using strict responseSchema with responseMimeType: application/json
• Using models like gemini-2.5-flash and gemini-2.5-pro

Critical requirement:

We strongly prefer NOT to chunk the PDF into smaller pieces.
The client wants:

• A single upload
• One single model call
• One single JSON array containing all issues
• No manual or automatic chunk splitting if possible

Chunking creates other difficulties (duplicate objects across chunks, page offsets, merging logic) which we’re trying to avoid unless absolutely necessary.

Client needs:

They want to upload any inspection report (even 300+ pages), and get reliable full-document structured JSON from Vertex AI in one go.

Looking for solutions / advice:

• Is there a Gemini/Vertex model with larger multimodal context than Flash/Pro?
• Any recommended architecture to process VERY large PDFs without chunking?
• Any way to handle image-heavy multi-page reports using Vision models efficiently?
• Is this a known limitation of Gemini’s multimodal window?
• Any roadmap for long-context Vision models on Vertex AI?

Thanks, any insights would be super helpful!