RAG as a service with parsing, indexing and checked citations built in
Building retrieval augmented generation yourself means parsers for five file formats, OCR for scans, chunking, an index, prompt rules, citation tracking and the part everyone skips: proving that the cited text is really on the cited page. ChatDocument runs all of it as a hosted service. You send documents and questions through a REST API or embed a widget, and you get answers with document, page and quoted passage, already verified.
Three endpoints for a RAG platform
- Upload: send a file or a URL into a space; PDF (scans read with OCR), Word (DOCX), Excel (XLSX, CSV), PowerPoint (PPTX), plain text, Markdown and web pages
- Ask: send a question for a space and receive the answer with citations (document name, page or slide, sheet and cell range for spreadsheets, quoted text)
- Extract: send a list of columns and get one row per document, each cell linked to its page
Spaces can also be filled from synced Google Drive, SharePoint and OneDrive or Dropbox folders, so your app reads documents your team already maintains.
The pipeline you do not have to maintain
- File parsing per format, with page, slide and cell positions kept
- OCR for scanned PDFs
- Chunking, ranking and an index per space
- Answer generation restricted to retrieved passages
- Citation check in code: each cited passage must exist on its page and contain the quoted text, or the sentence is dropped
- A defined response when nothing in the space answers the question, with the closest passages listed
- Deletion of documents and index on request or with the space
Grounded by design, not by prompt
Many RAG stacks tell the model to stay on the documents and hope it listens. ChatDocument enforces it after generation: answers come only from the documents in the selected space, and unsupported sentences never reach your users. A generic chatbot fills gaps from general knowledge, and that is exactly the failure that makes customers distrust an AI feature. Your application receives a clear "not in the documents" signal instead, which you can show, log or route to a person.
Limits a builder can plan around
API questions count toward the plan, the same as questions asked in the app. One question is one cited answer, one extracted row (one document) is 1 question, a comparison is 3. New pages count once per billing period, scanned pages count as 3. Deleting a document does not give pages or questions back.
| Business | Enterprise | |
|---|---|---|
| Price billed yearly | 249 USD per month (2,988 USD per year) | 745 USD per month (8,940 USD per year) |
| Price billed monthly | 499 USD per month | 1,490 USD per month |
| Questions per month, pooled | 7,000 | 20,000 |
| Questions per day | 800 | 2,000 |
| New pages per month | 30,000 | 100,000 |
| Max file size | 200 MB | 500 MB |
| Extraction per run | 250 documents | 1,000 documents |
| Seats | 15 | 50 |
| Extras | Widget, API, roles, audit log | Widget, API, roles, audit log, SSO (SAML), SLA 99.9, invoice billing |
Yearly billing saves 50%. Full plan table on chat with PDF pricing.
FAQ
Do I need my own model or vector database
No. Retrieval, generation and the citation check run inside ChatDocument. Content is processed by AI model providers bound by contract not to train on your content.
Can each customer of mine have separate documents
Use one space per customer. Each API call targets one space, so answers never mix documents across spaces.
Which plans include the API
Business and Enterprise. Starter and Team use the same engine in the web app.
What does the API return when there is no answer
A response that states the documents do not contain an answer, with the closest passages and their pages.
Ship grounded answers this week
Create the account, open a Business space and make your first upload call.