feat: Support PDFs as Model-Independent Context

[JamesYeh23]

Feature: Support PDFs as Model-Independent Context

Problem

PDF files are currently not supported as uploaded context, partly because not all models can read PDFs directly.

Some models support PDF input, while others only support text-based formats. If a PDF is added directly as context and the user later switches to a model without PDF support, the context may no longer be usable.

This creates several problems:

• PDF context cannot be reliably shared across different models.
• Switching models may cause the context to fail.
• Users may need to manually convert PDFs into text or Markdown.
• The original PDF is still valuable for source verification and visual reference.

Proposed Solution

When a user uploads a PDF as context:

1. Preserve and store the original PDF file.
2. Extract the readable content from the PDF.
3. Save the extracted content as a Markdown file associated with the original PDF.
4. Use the Markdown file as the primary context for model interactions.
5. Keep a reference to the original PDF.
6. Direct the user to the original PDF when they ask about page layout, images, tables, visual details, or source verification.

For example:

document.pdf
document.md

[fasuizu-br]

Affiliation: I am writing on behalf of BRAINIALL, which operates a metered hosted text-layer PDF-to-Markdown Actor and does not provide OCR. It is not PDF storage, source verification, visual layout, image or table inspection, a context model, or a Cometline integration. Before explicit maintainer opt-in I will not call a provider, upload any document, open a PR, or create a fixture; after opt-in I would use only one generated rights-cleared two-page text-layer PDF plus one image-only control. Since #96 leaves the extraction backend open while requiring model-independent Markdown and preservation of the original PDF, would one no-PR boundary receipt be useful solely to record fixture hashes, page-range request, Markdown or explicit empty behavior, status and MIME, elapsed time, public route and price, and observed charge without claiming OCR, layout fidelity, visual verification, storage, or integration compatibility? If there is no explicit maintainer opt-in within 7 days, I will treat this as killed and will not follow up.