Claude Files API, PDFs and images: limits and block types

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. Your document pipeline uploads scanned invoice images through the Files API and references them from later requests. You reference an uploaded PNG with a document content block and the request is rejected. What change would most effectively address this issue?

   Right. In the game that’s Renown for your Prompt Engineering, Structured Output and Retrieval building.

   Not that one: D is right. In the game your Prompt Engineering, Structured Output and Retrieval building would lose a little Renown, and this question would come back later.

   Your answerA.Re-upload the PNG as base64 inside a text block and describe its media type in the prompt.A text block carries prose. Images are sent either as an image content block with base64 data or as an image block referencing a file_id, which is the whole point of uploading once.Your answerB.Keep the document block and add media_type: "image/png" beside the file_id.The mismatch is in the block type, not a missing annotation. A document block referencing an image file is rejected however it is labelled; the correct reference is an image block.Your answerC.Convert the PNG to a single-page PDF so the document block matches an accepted type.Converting a PNG to PDF adds a processing step to work around a block-type error. The API already accepts images directly through an image block referencing the same uploaded file.The answerD.Reference the uploaded PNG with an image content block, so the block matches the MIME type.When a file uploaded through the Files API is referenced by file_id, the content-block type must match the file's MIME type: a document block for PDF and text files, and an image block for images.

   Show the full explanation and doc quotes

   A PNG therefore has to be referenced with an image block.

   Written from platform.claude.com/docs/en/build-with-claude/files

2. Your documentation team hosts its product manuals as PDFs on a public CDN. You want Claude Sonnet 5 (claude-sonnet-5) to answer questions about a manual without encoding the file into the request body and without storing a copy in Anthropic's file storage first. Which approach should you take?

   Right. In the game that’s Renown for your Prompt Engineering, Structured Output and Retrieval building.

   Not that one: A is right. In the game your Prompt Engineering, Structured Output and Retrieval building would lose a little Renown, and this question would come back later.

   The answerA.Send a document block whose source uses type: "url" with the link.A document content block accepts three source types: base64, url and a Files API file_id. The URL source points the API at a publicly reachable PDF, which is fetched for the request, so neither client-side base64 encoding nor an upload step is needed.Your answerB.Fetch the PDF in application code and inline it as base64 document data.Base64 works, but it forces the application to download and encode the manual and it inflates the request payload against the 32 MB limit. The URL source removes both steps for a file that is already publicly hosted.Your answerC.Upload the PDF through the Files API and reference the file_id it returns.The Files API is the right answer when a file is used repeatedly and should be stored once, but it means uploading a copy and managing its lifecycle. The requirement here is explicitly to avoid that.Your answerD.Paste the CDN link into a text block so the model retrieves the page itself.Text is not a retrieval instruction. Without a fetch tool, a link written in a text block gives the model the characters of the address and nothing about the document behind it.

   Written from platform.claude.com/docs/en/build-with-claude/pdf-support

3. Your pipeline uploads each night's PDFs through the Files API with a 24-hour expiry and reuses the ids the next morning. A reconciliation job lists the workspace's files, finds the ids present, and then every Messages request that references them fails before any tokens are generated. What is the most likely cause?

   Right. In the game that’s Renown for your Prompt Engineering, Structured Output and Retrieval building.

   Not that one: B is right. In the game your Prompt Engineering, Structured Output and Retrieval building would lose a little Renown, and this question would come back later.

   Your answerA.Listing a file refreshes its metadata but revokes its reference for the current day.Listing files and retrieving metadata are read-only operations, to the point that they are not recorded in the Compliance API activity feed. They cannot revoke a reference.The answerB.The files have passed expires_at, and expired files remain listed for 30 days.When a file reaches its expires_at, a Messages request that references it fails before inference, but its metadata remains readable for up to 30 days and it continues to appear in list responses during that window.Your answerC.The workspace has exceeded the 1 TB storage limit, which blocks new references.Exceeding the 1 TB organization storage limit produces a storage-limit error on upload. It does not cause an otherwise valid reference to fail at inference time.Your answerD.Expired files must be re-uploaded with a longer expires_in_seconds to be listed.Expiration is set once at upload with expires_in_seconds and cannot be changed. Uploading the content again produces a new file with a new id rather than extending the old one.

   Show the full explanation and doc quotes

   A job that treats presence in the listing as proof of usability will keep referencing dead ids.

   Written from platform.claude.com/docs/en/build-with-claude/files