Beginner
Project Knowledge Base
The knowledge base is where you store reference documents that Claude should have access to in every conversation within a Project. Upload once, reference always — without re-uploading in each session.
What to Upload
The knowledge base works best for reference material that Claude needs to consult repeatedly:
Good candidates for the knowledge base
- Style guides and brand voice documents
- Product specifications or feature documentation
- Client briefs, requirements, or contracts
- Reference codebases or API documentation
- Research papers or source material
- Company templates and standards
- Your own prior work (Claude can learn your patterns)
Not ideal for the knowledge base
- Very large corpora (hundreds of documents — context becomes diluted)
- Highly dynamic content that changes daily
- Sensitive credentials, passwords, or tokens
- Content only relevant to a single conversation (attach to that conversation instead)
Supported File Types
| Format | Notes |
|---|---|
| Text is extracted; scanned/image-only PDFs may have limited extraction | |
| DOCX / DOC | Word documents — text content extracted |
| TXT, Markdown | Plain text and Markdown files — fully supported |
| CSV | Tabular data — Claude can query and reason about rows and columns |
| Code files (.py, .ts, .js, etc.) | Treated as plain text — Claude reads and reasons about code |
How Claude Uses the Knowledge Base
The knowledge base is not a verbatim lookup — Claude uses it as additional context in conversations. How Claude accesses it:
- Automatic retrieval: Claude intelligently surfaces relevant parts of your knowledge base documents when they are relevant to the conversation, without you needing to reference them explicitly.
- Explicit reference: You can refer to a document by name: "According to our style guide..." or "Based on the project spec you have access to..."
- Full document available: For documents within the context budget, Claude can access the full text, not just retrieved chunks. This is different from a traditional RAG system.
Context budget affects large knowledge bases
If your total knowledge base exceeds the context window, Claude cannot load everything at once. It will retrieve the most relevant portions. For very large knowledge bases (thousands of pages), consider using a dedicated RAG pipeline instead.
Updating Knowledge Base Files
- Replacing a file: Delete the old file from the knowledge base and upload the new version. There is no in-place update — you replace, not patch.
- Appending content: You can upload multiple files with related content. Claude treats them as separate documents. Name them clearly so you can reference them by name in prompts.
- Version management: If you version your documents (e.g. "style-guide-v2.pdf" vs "style-guide-v3.pdf"), make sure to remove old versions when they're superseded to avoid Claude referencing outdated content.
Prompting Strategies for the Knowledge Base
- Reference by filename: "Using the product spec document in this Project, what are the requirements for the authentication module?"
- Cross-reference multiple files: "Compare the requirements in spec-v1.pdf to spec-v2.pdf — what changed?"
- Ask what's in the knowledge base: "What documents do you have access to in this Project? Summarise each one in a sentence." Useful for confirming what Claude can see.
- Explicitly ground responses: "Answer only based on the documents in this Project — do not use your general knowledge." Reduces hallucination risk for domain-specific work.
Checklist: Do You Understand This?
- Upload reference documents once — they're available in every conversation without re-uploading
- Supported formats: PDF, DOCX, TXT, Markdown, CSV, code files
- Large knowledge bases may exceed the context window — Claude retrieves the most relevant portions
- To update a file, delete the old version and upload the new one — no in-place update
- Reference files by name in prompts for more targeted retrieval; ask Claude to ground answers in knowledge base docs to reduce hallucination