Use case
How to train a chatbot on a PDF
Last updated by Venu Gopal Reddy Muvva, Founder, Surge Software Solutions
Who this is for
This is for you if your answers live in documents, not on a website. You might have a price list, a policy PDF, a product manual, a course brochure or a long FAQ in Word. Customers ask about it every day and you copy and paste the same parts. You want a chatbot that reads the document and answers from it, and doesn't guess.
How to do it in CustomerReplies
- Sign in. Go to /sign-in and enter your email. We send a 6-digit code. In onboarding, the website address is optional.
- Upload your file. On the Sources page, add your PDF, Word (.docx), TXT, Markdown, HTML or CSV file. Each file can be up to 8 MB.
- Add Q&A for gaps. If customers ask things the file doesn't cover, add Q&A pairs, or import them from a CSV file.
- Test in the Playground. Ask questions the file answers. Check the citation. Correct any wrong answer once.
- Go live. Copy the script tag from the Widget page into your site, or share a link to a full-page chat.
Get your file ready
The agent reads the text in your file. Files with real, selectable text work best. If your PDF is a scan or mostly pictures, check the answers in the Playground. If they are poor, paste the text in as a text source instead.
Clear headings help. Short sections with one topic each are easier for the agent to find and cite.
Keeping answers right
The agent answers only from your sources. The fact guard marks answers with prices or phone numbers that are not in your sources as low confidence, and the link guard removes links that are not in them.
When the document changes, upload the new version on the Sources page and remove the old one. If you keep the document in Google Docs or a public Notion page instead, the agent can re-sync it daily or weekly.
What you pay
The first 3 months are free. After that you pay only for what you use, at our cost — no markup. Monthly plans are coming soon.
When CustomerReplies is not the right tool
CustomerReplies is not the right tool if your files are larger than 8 MB each and you can't split them. It is also not right if you need answers in Hindi or another language; it understands them but answers in English. Private Google Drive folders are in Beta, so share files by link or upload them for now.
Questions
Can I upload more than one file?
Yes. Add as many files as you need on the Sources page. The agent answers from all of them and cites the one it used.
What about tables in my PDF?
The agent reads the text in them. Simple tables usually work. Check them in the Playground, or upload the table as a CSV file.
Do you train AI models on my files?
No. We don't train AI models on your data.
How much does it cost?
The first 3 months are free. After that you pay only for what you use, at our cost — no markup. Monthly plans are coming soon.