PDF / documents

PDFs and images can be converted to markdown before entering the context — once converted it's ordinary text: searchable, quotable, compressible, and it doesn't run up a vision model's bill.

Decoupled from the active provider

Document parsing goes through GLM-OCR, independently of the model you're working with:

Working on DeepSeek + a GLM key configured ⟶ PDFs still parse

What it needs is a GLM key (env ZHIPUAI_API_KEY, or providers.glm.apiKey in settings). It does not require switching your active provider to GLM. Without a GLM key it reports a missing key explicitly rather than failing silently.

Output

You get markdown plus a page count. The converted content enters the context as ordinary text — you can have it Grep the document, quote a specific section, cross-reference another file, none of which is possible when a PDF is treated as a picture.

Timeouts and errors

Situation	What you see
Over 180 seconds	"document parsing timed out"
Non-2xx from the server	"document parsing failed: HTTP <status>"
Unexpected response shape	"document parsing failed: no md_results in response"

All three are explicit errors — you never get an empty document that looks like a successful parse.

It honours https_proxy / HTTPS_PROXY / http_proxy / HTTP_PROXY when a proxy is needed.

Which one to use

• Screenshots, charts, UI problems → Image vision, let the model look
• PDFs, scans, long documents with layout → document parsing, convert to markdown

The deciding question is whether you'll want to do text operations on it afterwards. If you need to search, quote, or cross-reference, convert to markdown first.

---

Related: Image vision · Providers