You can build a chatbot that answers from PDFs and other documents by uploading your current files, waiting for processing, and testing its answers against the original documents. In Agentkit, start at Sources > Files. Supported uploads are PDF, DOCX, XLSX, PPTX, TXT and Markdown.
Document chat retrieves relevant content from a searchable knowledge base and uses it to compose an answer. It can help visitors find a warranty rule or employees find a procedure. It still needs answer checks. A fluent response does not prove that a table, exception or date survived extraction correctly.
Prepare the documents
Choose the current version of each document. Remove obsolete versions that contradict it. Decide who should see the information before publishing an agent trained on internal material.
| Upload format | Extension | What to check |
|---|---|---|
| Dates, exceptions, scanned pages and tables | ||
| Word | .docx | Headings and the rules they introduce |
| Excel | .xlsx | Sheet names, units and relationships between cells |
| PowerPoint | .pptx | Slide text and information conveyed only visually |
| Plain text | .txt | Enough context to interpret the answer |
| Markdown | .md | Section structure and current links |
CSV is not an advertised file-upload format. Export the relevant data to a supported format and check the result. Changing a filename extension is not conversion.
Scanned PDFs and complex layouts need particular care. Agentkit's processing includes a fallback extraction path for documents that local conversion cannot handle. That is not a guarantee that every scan, diagram or table will be read correctly. If extraction fails or the answer misses important details, use a clearer text-based version and test again.
Upload and check processing
- Open your agent and go to Sources > Files.
- Upload up to 20 supported files in one batch. Each file must fit the 50MB upload limit.
- Wait for processing to finish. Investigate failures before relying on those files for answers.
- Check your remaining plan storage. It counts processed content, so the original PDF's file size does not tell you exactly how much knowledge-base storage it will use.
See the document source documentation for source management. Upload limits and plan storage are separate constraints. A file can fit the upload limit while its processed content exceeds the storage available to your agent.
Test a specific answer in Playground
Choose a question with an unambiguous answer in the original file. Record the expected answer and the relevant section, then ask the agent in Playground.
For example, a warranty manual might allow two years of coverage with an exception for consumables. Check that the answer includes the period and the exception. This is an illustrative test, not a measured product result.
Repeat with a table-based question, a follow-up and a question the documents cannot answer. Check dates and units, not just whether the response sounds plausible. Retrieval helps ground answers; it does not guarantee accuracy or prevent every invented detail.
Add a Q&A source for an important explanation missing from your files. It does not guarantee that the agent will repeat the answer verbatim or ignore conflicting sources. Correct the obsolete source as well.
Combine files with website content
Use files for manuals and information that is not online. Add current public help pages for broader coverage. The website training guide explains crawling, sitemap imports and individual links.
Keep public and internal knowledge separate when they need different access. A document upload does not make sensitive information suitable for a public widget. Static files are also a poor source for rapidly changing inventory or live account information.
Replace outdated versions and retest
Uploaded files are snapshots. Website auto-retraining does not replace a document when its original changes. Remove the obsolete file, upload the current version, wait for processing and repeat the questions affected by the change.
Watch for a correct general answer that uses an old date or omits a new exception. Those small differences can change the advice a customer receives.
Publish a tested document chatbot
Once the answers pass your checks, open Channels and copy the installation code generated for your agent. Follow the embedding guide, then test the installed widget with the same questions.
You can also add website sources or connect a human escalation action for questions that need your team. A connected support destination and a paid plan are required for that action.
Create an Agentkit agent to try the workflow with one current document. Expand the source set after you have checked the questions it should answer.
