Add Knowledge to a GPT: Limits, Formats, Testing
Your GPT’s memory is only as good as its notes. Feed it clean, structured docs and it’ll quote like a star student.
Last updated:
TL;DR
- Upload text‑first files with clear headings. Avoid image‑only PDFs.
- Use separate, themed files (≤ 20 files total) with descriptive names.
- Keep each file ≤ 512MB and ≤ ~2,000,000 tokens; CSVs ~≤50MB; images ≤20MB.
- In Instructions, tell the GPT to prefer Knowledge and quote sources.
- Test with a “golden questions” list; fix misses by restructuring files, not just adding more.
Where to upload
In the GPT editor under Build / Configure → Knowledge. Drag files or pick from your computer/cloud. The GPT indexes your text into chunks for retrieval.
Recommended formats
- Markdown (.md) or Plain text (.txt) — cleanest extraction.
- PDF — okay if single‑column; avoid complex layouts and scanned images.
- CSV — great for tables, but keep sizes modest; use a header row.
Current limits (practical table)
| Item | Limit | Notes |
|---|---|---|
| Files attached to one GPT | Up to 20 files | Attach in GPT editor → Knowledge. |
| Per‑file size (most types) | ≤ 512MB | Hard cap. |
| Per‑file token cap (text/docs) | ~2,000,000 tokens | Text is chunked & embedded; very large docs can slow indexing. |
| CSV/Spreadsheet size | ~≤ 50MB | Practical ceiling varies by row size. |
| Images | ≤ 20MB each | Only text in files is used for Knowledge retrieval; images aren’t indexed as Knowledge text. |
| End‑user storage cap | ~10GB | Total uploads per end‑user across features. |
Limits change — check for updates before quoting hard numbers.
File naming & structure that actually works
Name files by topic
/knowledge/ 01-product-overview.md 02-pricing-policy.md 03-api-auth.md 04-api-endpoints.md 05-troubleshooting-faq.md
Use Markdown scaffolding
# API Authentication ## Overview Short summary. ## Steps 1. Obtain API key from ... 2. Send header: `Authorization: Bearer <key>` ## Error Codes - 401: Invalid key — rotate via ... ## Examples ```bash curl -H "Authorization: Bearer $KEY" https://api.example.com/ping ```
Instruction block to force Knowledge usage
Use the uploaded Knowledge as the primary source. When answering, first retrieve and quote the most relevant section(s) with headings. If Knowledge lacks the answer, say so briefly, then continue with general expertise. When relevant, show file name and heading like: [03-api-auth.md › Error Codes]. Avoid browsing unless the user asks for live info.
Testing rig (“golden questions”)
- Create 10–15 real questions users ask (mix easy/hard/edge cases).
- Ask them in a fresh chat with your GPT; record pass/fail and notes.
- For fails: rewrite the relevant section and split oversized files.
- Re‑run the same questions after each change until you reach ≥90% pass.
Pro tip: add an internal doc “Changelog & Decisions” and keep it in Knowledge so the GPT can explain why choices were made.
Troubleshooting
- “GPT ignores my files” → Ensure content is text‑first; convert scans to text; reduce multi‑column layouts.
- “It hallucinates despite Knowledge” → Add negative rules in Instructions and require quotes from files.
- Poor recall → Split mega‑files by topic; add headings and repeated keywords users actually type.
- Data analysis tasks fail → Enable the GPT’s data/code tool and keep CSVs tidy with headers.
FAQ
How many files can I attach?
Up to 20 files per GPT.
Are images parsed?
Only text is used for Knowledge retrieval. Include alt text or captions in the file body if needed.
What file sizes are supported?
Most files ≤512MB; text/docs also have a ~2M token cap; CSVs ~≤50MB; images ≤20MB.
Why won’t it use my PDF?
Multi‑column or scanned PDFs don’t extract well. Convert to text/Markdown or simplify layout.