"The PDF I uploaded yesterday isn't being referenced today." If you use Claude Projects daily, you've probably hit this exact wall. There's no error, no warning — Claude just calmly replies "I don't have information about that." That silence is what makes it so frustrating.
This guide walks through the failure modes I've personally hit while using Claude Projects in production work, organized by where the problem actually lives. Once you can tell whether the issue is the file, the index, or your custom instructions, recovery becomes a five-minute exercise instead of an afternoon-long debugging session.
Suspect the index, not the file, first
When you upload a file to a Claude Project, it isn't immediately searchable. The platform chunks the document, generates embeddings, and writes them to a searchable index. That pipeline takes time.
In my experience, a clean text-based PDF under 30 pages is searchable within a minute or so. A 100-page report packed with images can take five to ten minutes. A spreadsheet with hundreds of rows can take longer.
Symptom: I asked right after upload and got "no relevant information" Wait five minutes, then start a new conversation. The same conversation tends to inherit its earlier "I don't know" judgment, even after the index updates.
Symptom: Hours later, still no recognition Now you can suspect the file itself. Move to the next section.
Four checks on the file itself
1. Is it a scanned PDF?
This is the single most common cause. A PDF exported from Word has a real text layer. A PDF created by scanning paper is a stack of images. Claude Projects does not run OCR, so a scanned PDF is invisible to knowledge search.
The quickest test: open the PDF and try to select text with your cursor. If you can't, it's a scanned PDF. Run it through Acrobat OCR or Preview's text-recognition feature, then re-upload the new version.
2. Is the file too close to the size limit?
The per-file limit is 30 MB, but anything above 20 MB starts to fail more often. If you're near the ceiling, split the file by chapter or section before uploading.
3. Is it password-protected or restricted?
PDFs with "no copy" or "no print" restrictions sometimes block text extraction. Remove the password, save a fresh copy, and re-upload.
4. Are you fighting a character-encoding issue?
This one is sneaky. Japanese, Korean, or other non-ASCII docx files that have bounced between Mac and Windows can have encoding artifacts that confuse Claude's text extraction. Copy the content into a plain .txt file as UTF-8, then upload that. The symptom often disappears entirely.
When files seem to be "disappearing"
"I had eight files yesterday, now I see six" is another common report. The usual culprits:
Cause A: Project storage cap
Free and Pro plans have a per-project knowledge cap. Once you cross it, new uploads appear to succeed but older files quietly fall out of search relevance. The fix is to delete material you've already merged into other documents, freeing capacity for what matters.
Cause B: You're signed into a different account
People juggling work and personal accounts hit this constantly. Browser session expiry can silently swap you to a different workspace where the project doesn't exist. Check the account name in the corner before you panic.
Cause C: The project got auto-archived
Long-dormant projects can get pushed off the main list even without an explicit archive action. Check the "Archived" filter in your project list.
Knowledge is indexed, but Claude refuses to use it
Different problem, different fix. Files are present, search would find them — yet Claude answers from generic knowledge instead.
Each project has its own custom instructions (system prompt). If yours says something like "always answer concisely" or "prioritize your training knowledge," the model often deprioritizes knowledge lookup.
The workaround I keep in every production project is a single line at the top of custom instructions:
Before answering, always search the project knowledge first.
If anything in the knowledge applies, quote and prioritize it over general knowledge.
That one sentence has more impact than any other tuning I've done. Without an explicit nudge, Claude tends to be conservative about when to actually run a knowledge search.
When nothing works
If you've cleared every check above and the project still misbehaves, try these in order.
- Spin up a fresh project and copy only the knowledge over. Project state can corrupt in rare cases. A clean container often resolves it instantly.
- Start a brand new conversation. Long conversations crowd the context window and hurt knowledge recall.
- Test in a different browser. Some Chrome extensions (especially ad blockers) interfere with the upload API. Try an incognito window or a different browser.
- Document repro steps before contacting support. Note which file, which question, what time. Anthropic support can act much faster on a tight repro than a vague "it's broken."
What to try first
Run the four file-side checks before doing anything else. In my experience, nine times out of ten the answer is either "scanned PDF" or "encoding artifact." If knowledge is being ignored, add the explicit "search first" line to your custom instructions — it changes the project's behavior more than any other setting I know of.