Claude AI and PDF Reading in 2025: Capabilities, Limits, and Workflows

Graziano Stefanelli
Oct 14
4 min read

Claude can read, search, and cite long PDFs across the web app, Projects, the API, and partner platforms (Vertex AI, Amazon Bedrock). In practice, you’ll mix three building blocks: file upload, grounded Q&A with citations, and context-window management. Here’s how it works now—and how to avoid the common pitfalls.

·····

.....

What Claude actually does with a PDF.

When you upload a PDF, Claude extracts the text layer and can also reason over page images (for layouts, figures, or scanned pages). In API responses, citations can reference page locations—so answers point back to the exact PDF page(s) used. That makes audit trails and compliance reviews much easier.

In the Claude.ai web app, you can drag-and-drop PDFs directly into a chat or a Project (a focused workspace with its own knowledge base). Project knowledge becomes persistent context for all chats in that workspace.

·····

.....

File types, sizes, and how many you can add.

Supported: PDFs plus common office/text formats (DOCX, TXT, CSV, RTF, EPUB, HTML) are accepted across chat and Projects. Typical limits: up to 20 files per chat, ~30 MB per file; Projects accept many files but your usable context still caps out at the model’s window. Treat size and tokens as separate constraints.

If you’re using the API, the Files API lets you upload once and reuse files across calls—use the 2025 beta header to enable it. This avoids re-sending large PDFs every request.

·····

.....

Context windows: how much a single run can “see.”

Claude’s standard docs describe ~200k tokens of context (shared by your prompt, the PDF text it pulls in, and the model’s output). Enterprise and newer Sonnet releases are rolling out larger windows (up to 1M tokens) for high-tier API users, which materially changes what you can load at once (think multi-hundred-page filings). Always check your account’s current limit before planning a workflow.

Rule of thumb: don’t aim to stuff an entire repository of PDFs into a single turn; curate the smallest set that answers the question, then chain turns (or use Projects) to keep the conversation grounded.

·····

.....

Grounded Q&A with page-level citations.

Use the Citations feature to require Claude to justify answers with explicit references. In the API, enabling citations yields structured references—PDFs return page-based locations—which you can render in your UI or export to reports. This is available on the Anthropic API (and via Bedrock) and is the recommended path for regulated workflows.

Prompt pattern:

“Answer only from the attached PDFs. Provide page-level citations for each claim. If evidence is missing, say ‘insufficient evidence’.”

·····

.....

Web app vs. Projects vs. API: when to use which.

Web app (Chat): fastest path for a single report—drop in a PDF, ask questions, export answers with references. Great for ad-hoc analysis.
Projects: persistent knowledge base of PDFs and docs for an ongoing initiative (case file, deal room, policy library). Chats in that Project can reuse the same sources without re-uploading.
API + Files API: productionize the flow (batch ingestion, job queues, verifiable citations, role-based access). Pair with your storage and add guardrails in prompts.

You can also call Claude’s PDF abilities on Amazon Bedrock or Vertex AI if your stack lives there; code samples show end-to-end “send a document, get analyzed output” flows.

·····

.....

Practical limits and gotchas.

Password-protected PDFs: remove passwords before upload—these are not processed. (Common limitation across apps/APIs.)
Scanned PDFs: Claude can reason over page images, but quality depends on the scan. If you own the source, extract text first for best results; otherwise expect more tokens per page due to vision analysis.
Token vs. file size: a 10 MB PDF with dense text can still overrun context; a 20 MB brochure might not. Plan by tokens, not megabytes.
Many files at once: the chat UI allows ~20 per turn; in production, upload via the Files API and reference only what you need in each request.

·····

.....

Build: a clean pipeline for reliable PDF answers.

Ingest PDFs to the Files API (store IDs).
Select the minimal set of file IDs for a given query (avoid context bloat).
Prompt with a strict rubric: scope, audience, answer length, citations required.
Validate: check citations return the right pages; fail the run if no citations appear.
Iterate: store feedback, refine prompts, and (optionally) index PDFs externally if you outgrow single-run windows.

·····

.....

What’s new in late 2025.

Create & edit files: Claude can now generate PDFs, slides, spreadsheets, and docs directly in the app/desktop—useful for “summarize → export” or “draft a brief with exhibits” loops.
Bigger windows: staged rollout of ~1M-token contexts for Sonnet tiers expands single-turn document capacity dramatically—plan for it, but feature-check your account first.

·····

.....

Troubleshooting checklist.

The model “forgets” parts of the PDF → you exceeded your token window. Reduce files/pages or switch to a larger context tier.
Upload succeeds but citations don’t show pages → confirm citations enabled and that you’re passing files through the Messages API/Files API correctly.
Slow/expensive runs on scanned PDFs → pre-OCR with your pipeline or split the doc; image reasoning consumes more tokens.

·····

.....

DATA STUDIOS

..... [datastudios.org]