Product manuals
“How do I reset it to factory settings?”
Step-by-step content retrieves particularly well because each procedure is self-contained.
PDF training
Train a chatbot on your PDFs, and the manual nobody reads starts answering
Upload the PDF. CustomerBot extracts the text — including from image-heavy pages — splits it into passages, indexes it, and answers from it with the document named. A download link becomes an answer.
Two things are called a PDF chatbot
They sound identical and solve opposite problems. Worth thirty seconds now rather than a disappointed signup later.
Not this
Reading a document yourself
You upload a research paper, a contract or a report and ask it questions privately, to save yourself the reading. ChatPDF, Adobe Acrobat’s assistant and SciSpace do this well.
This page
Making your documents answer other people
Your manuals, policies and price lists become answers your customers get on your own website, in the chat widget, without downloading anything. That is this page.
Four documents worth uploading today
Price and fee tables
“What does the mid tier cost annually?”
Tabular data is extracted as text, so the numbers stay attached to their row labels.
Terms and policies
“Can I cancel in the first fourteen days?”
The clause is quoted back in plain language, with the document named.
Spec sheets
“Is it rated for outdoor use?”
Dense reference data that nobody wants to read but everybody wants to query.
What happens to the file after you drop it in
- 01Upload
Drop the file into the training screen. It joins the same knowledge base as your crawled pages, text snippets and Q&A pairs — one index, not four.
- 02Extract
The text layer is pulled out. Pages that are mostly image — scanned tables, diagrams, screenshots of a form — go through image extraction so their content is recovered rather than silently dropped.
- 03Chunk
The document is split into overlapping passages. A 90-page manual is not one blob; it becomes hundreds of retrievable pieces, each small enough to be specific.
- 04Embed and index
Each passage becomes a vector in the same index as everything else, so a question can be answered from a PDF and a web page in the same reply.
- 05Answer with the source
The reply names the document it drew on, so anyone can check it against the original.
What extracts well, and what does not
PDF is a printing format, not a data format. Being honest about that saves you from wondering why one document works beautifully and another does not.
| Extracts cleanly | Expect trouble |
|---|---|
| A born-digital PDF exported from Word, InDesign or a CMS | A photo of a page taken on a phone at an angle |
| Selectable text you can highlight in a PDF reader | A flattened scan with no text layer at all |
| Tables with clear headers and consistent columns | Multi-column magazine layouts that interleave when read linearly |
| One document per subject | A 400-page compendium of everything you have ever published |
You do not have to build the pipeline
Search this topic and most of what you find is a tutorial: LangChain, a vector database, an embedding model, a weekend. That is a genuinely good way to learn how retrieval works, and a poor way to get a manual answering customer questions by Thursday.
Everything on the right is already done. Build it yourself when the learning is the point, or when your requirements are strange enough to need it.
What the retrieval actually doesChunking strategy
Pick a splitter, tune overlap, re-tune when answers straddle sections.
Embeddings and a vector store
Choose a model, run Pinecone, FAISS or pgvector, keep it upgraded.
Text extraction
Handle the born-digital case, then handle scans separately.
Re-indexing on change
Write the job that notices version 4 replaced version 3.
A widget and a backend
Build the chat UI, host the API, keep model keys out of the browser.
One index, four kinds of source
PDFs are not a separate mode. They sit in the same knowledge base as crawled URLs, pasted text and Q&A pairs, which means a single answer can draw on your terms PDF and your pricing page at once — and cite both.
PDF questions
What is a PDF chatbot?
A PDF chatbot is an assistant that reads your documents and answers questions about them in plain language instead of making somebody search the file. The term covers two fairly different products: tools that let you upload a document and interrogate it privately for your own reading, and tools that turn your documents into answers your customers get on your website. CustomerBot is the second kind — you train it on your manuals and policies once, and your visitors get the answers in the chat widget.
Is this the same as ChatPDF or Adobe’s PDF assistant?
No, and if that is what you want, those are the right tools. They are built for one person reading one document: upload, ask, done. CustomerBot is for the other direction — your documents, trained once, answering your customers on your own site alongside your web pages, with lead capture and human handover attached. Different job, similar-sounding name.
Do I need to build a RAG pipeline, or use LangChain?
No. The retrieval-augmented pipeline everyone builds in those tutorials — extract, chunk, embed, store, retrieve — is what the product already is. There is no vector database to run, no embedding model to choose and no chunking strategy to tune. If you want to understand what is happening under the hood, the training guide walks through it.
Can visitors upload their own PDFs to it?
No. You train the assistant on your documents; visitors ask questions and get answers from them. It is a knowledge base, not a document-analysis tool, and that distinction is deliberate — it keeps answers grounded in content you control and approved.
What about Word documents or spreadsheets?
PDF is the supported upload format. In practice this is less limiting than it sounds, because exporting a Word or Pages document to PDF takes seconds and improves extraction. For anything else — a table you maintain in a spreadsheet, say — pasting the content as a text source works well and is easier to keep current.
What happens to scanned PDFs?
Image-heavy pages are put through extraction so text inside them can be recovered rather than skipped. Quality depends on the scan: a clean, straight, high-resolution scan does well, a crooked phone photo does not. If a document matters and the scan is poor, the fastest fix is to paste the text directly as a text source.
Can it answer from tables?
Yes, within reason. Tables are extracted as text, which preserves the relationship between a row label and its values for most well-formed tables. Very wide or deeply nested tables can lose structure — if a fee table is critical, it is worth checking a couple of answers against the original before you rely on it.
How large a PDF can I upload?
The practical limit is your assistant’s storage: 5 MB on the free plan and 40 MB on paid plans. Because documents are stored as extracted text rather than as the original file, that covers a great deal more paper than the numbers suggest.
Can I mix PDFs with website content?
Yes, and you generally should. A single assistant can be trained on crawled URLs, uploaded PDFs, pasted text and Q&A pairs at once. Retrieval draws from all of them, so a question can be answered from a policy PDF and a pricing page in the same reply.
What if I update the document?
Replace it in the training screen. The old passages go with it, so there is no risk of the assistant quoting version 3 after you have published version 4 — which is precisely the failure mode of leaving old PDFs on a website.
Does it answer in other languages?
Yes. It detects the language of the question and replies in it across 95+ languages, reading from the source document in whatever language that was written.
Upload the document you keep emailing people
Then ask it the question you get asked about that document. That is the whole test, and it takes two minutes.
Start free