3.2¶
A PDF the server cannot fetch can now reach it through a one-time upload
link: the user or the agent downloads the paper through their own library and
sends the file straight to the server, which converts it like any other paper.
NIST shorthand such as nist 800-53r5 now resolves to the right publication.
VLM enrichment works with newer OpenAI models such as gpt-6-luna. A new
setting passes model-specific request parameters.
HTTP deployments with a public base URL gain the upload tool and its route on
upgrade. See Upgrading.
Upgrading¶
If you run the server over HTTP with SCHOLAR_MCP_BASE_URL set, which every
OIDC deployment does, the upgrade adds the create_upload_link tool and an
upload route under /transfer/. Nothing needs configuring. The tool is a
write tool, so read-only mode hides it. Under stdio, or without a base URL,
the server registers neither.
If you use VLM enrichment, its requests now limit output with
max_completion_tokens in place of max_tokens. Check that your VLM endpoint
and model accept max_completion_tokens.
Clients: reconnect to pick up the new tool. When fetch_and_convert
serves a PDF the server already holds, pdf_source is now "cache". In 3.1 it
reported the first source the server would have tried.
State: the cache, downloaded PDFs and converted Markdown are kept.
Uploaded PDFs are stored in the same pdfs directory.
Security posture: HTTP deployments with a base URL gain an
unauthenticated upload route, guarded by single-use capability URLs as the
security model describes. Each link admits one PDF
into the cache under a file name the server fixed in advance. The bytes must
start with %PDF-. A valid PDF already stored there is never overwritten.
Every client that asks for the paper afterwards is served the uploaded copy.
PDFs from your own library¶
Since 3.1, a paper the server cannot fetch comes with access links into the user's library. On an institution's licensed network such a link often opens for the user while the server, on another network, still has no copy. #546 puts the gap this way: there was "no way to get that PDF to the server for conversion" unless agent and server shared a disk.
create_upload_link closes it
(#600):
- The agent calls
create_upload_linkwith the paper's identifier and gets a one-timeurl. - The user or the agent sends the PDF as the body of an HTTP
PUTto that URL. The file goes to the server directly and never enters the conversation. - The agent calls
fetch_and_convertwith the same identifier and gets the paged Markdown, as for an open-access paper.
When fetch_and_convert finds no PDF, its response now points the agent at
create_upload_link. A document that isn't a known paper can be uploaded
without an identifier: the response then carries a pdf_path for
convert_pdf_to_markdown. If a converted paper turns out to be the wrong
document, create_upload_link with replace set to true removes the stored
PDF and its Markdown before the new upload.
An uploaded PDF joins the shared cache, so later requests for the paper get it too; the maintainer decided this on #546, and there is no per-upload opt-out. No route embeds the PDF in a tool call either: #597 closed that as not planned, because a PDF read into the conversation takes over the agent's context.
The tool exists only on an HTTP deployment with SCHOLAR_MCP_BASE_URL set.
Link lifetime and upload size follow the SCHOLAR_MCP_TRANSFER_* settings in
the configuration reference; by default a
link lasts an hour and an upload may be up to 100 MiB. The
PDF conversion guide
walks through the steps, and Transfer links
covers the operator side.
NIST shorthand¶
get_standard and resolve_standard_identifier resolved nist 800-53r5 to
NIST SP 800, a different publication, while nist sp 800-53r5 worked
(#585). The form
without sp now resolves the same way, so nist 800-53r5 gives
NIST SP 800-53 Rev. 5 and nist 800-171r2 gives NIST SP 800-171 Rev. 2
(#599). NIST IR numbers
written without nistir, the NIST.SP.800-53r5 form and multi-part numbers
such as 800-63-3 are still open in
#598.
VLM enrichment on newer OpenAI models¶
With use_vlm=true, docling-serve asks the configured vision model to read
formulas and describe figures. With gpt-6-luna, OpenAI refused the
requests 3.1 sends, because max_tokens "is not supported with this
model"; they want max_completion_tokens instead
(#603). The requests
now send max_completion_tokens with the same limits
(#605).
A model can also restrict other parameters. gpt-6-luna accepts only its
default temperature of 1, while formula extraction sends a temperature of 0
(#606). The new
SCHOLAR_MCP_VLM_PARAMS takes a JSON object whose keys replace the defaults
in every VLM call
(#607). For such a model:
export SCHOLAR_MCP_VLM_PARAMS='{"temperature": 1}'
Left unset, the requests keep docling's defaults. A value that isn't a JSON object stops the server at startup with an error naming the variable. The VLM section of the PDF conversion guide covers the setup.