ChatPDF
Document ProcessingAsk questions about PDF documents and get answers grounded in their contents, with optional page-level citations. Agents add a PDF once and query it as many times as the workflow needs.
What This Integration Enables
ChatPDF is not a field extractor and treating it as one is the fastest way to be disappointed by it. There is no schema, no template and no typed output. What it does instead is answer a question about a document in the document's own terms, and optionally tell you which pages the answer came from. That makes it the right tool for the documents that do not have a fixed shape: contracts, policies, technical manuals, board packs, regulatory guidance. The ones where the useful question is "does this agreement allow assignment on a change of control" rather than "what is in field seven".
For agents that changes what the connector is for. A FlowRunner agent adds a PDF once, receives a source ID, and then queries it as many times as the flow needs, branching on what the document actually says rather than on metadata about it. ChatPDF is stateless between calls, so multi-turn questioning works by passing earlier turns back in. Every operation is synchronous, so there is no job to poll. The discipline this connector demands is provenance: an answer generated from a document is a claim about that document, and FlowRunner's job is to make sure a claim nobody can trace never becomes a decision nobody can defend.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Contract intake without reading the contract
A countersigned agreement is attached to an intake issue in Jira. The agent calls Upload PDF File, which works even though the attachment sits behind authentication, because FlowRunner fetches the bytes, verifies they really are a PDF, and posts them to ChatPDF itself. With the source ID in hand the agent works through the intake checklist one question at a time with Include Page References enabled: term and renewal, governing law, liability cap, assignment on change of control, termination for convenience. Each answer is written back onto the issue with the pages it came from, so the person who reviews it can jump straight to the clause rather than trusting the summary.
Screening inbound documents before routing them
Vendor security questionnaires, insurance certificates and subprocessor lists all arrive as PDFs in the same intake folder. The agent adds each one and asks a single screening question that decides the branch: what kind of document is this, and does it name an expiry date. Documents with an expiry inside the review window are routed to the owning team in Microsoft Teams with the date and its page. Documents that are not what the folder expected are held. The flow branches on what the document says, which is not something a filename or a MIME type can tell you.
A running conversation about one long document
An analyst working through a two-hundred-page regulatory guidance document asks a follow-up question that only makes sense in the context of the previous two. ChatPDF is stateless, so the agent passes the earlier turns back in the Messages list. It also keeps the thread short on purpose, because at most six messages are accepted per request and a conversation past roughly two and a half thousand tokens has its oldest messages dropped server-side without warning. The agent therefore restates the constraint that matters in the current question rather than relying on the model to still be holding it, and it records the full exchange in Notion where the human record is complete even when the model's is not.
Human-in-Loop Highlight
Include Page References is the single most important switch on this connector, because it is the difference between an answer and a citation. When it is on, ChatPDF returns bracketed page markers in the answer along with a structured list of the pages each claim was read from. When an answer comes back without any page reference attached, the agent has been handed a statement about a contract that it cannot point to anywhere in the contract. It may well be right. It is not checkable, and an uncheckable answer about a liability cap or an assignment clause is exactly the kind of thing that gets repeated in a summary, then in a memo, then in a negotiation.
So the agent stops there. It does not paraphrase the answer more carefully, and it does not try a differently worded question and take whatever comes back second. It escalates to counsel in Slack with the question it asked, the answer it received, the source PDF, and one line stating the problem: "ChatPDF answered that assignment is permitted on a change of control, but returned no page reference for it. I have not written this to the intake record. Can you confirm the clause, or should I flag the agreement for full review?" Counsel answers, and either the citation is supplied and the record is complete or the agreement goes to a person. This is what an automation exception looks like when the doubt is about provenance rather than about a number: the agent has done real work, and it hands over at exactly the point where its evidence runs out.
Agent Capabilities
4 actionsSources
2- Add PDF From URL Registers a PDF already hosted at a publicly reachable URL and returns the source ID every other operation needs. ChatPDF downloads and indexes it itself, so nothing transfers from FlowRunner. The URL must be reachable without authentication, and the document must be a PDF within ChatPDF's page and size limits. Preferred over uploading whenever the document already has a public link, because it avoids moving the file twice.
- Upload PDF File Uploads a PDF from any URL FlowRunner can reach, including private or signed links ChatPDF could not download on its own, and returns the source ID. The bytes are fetched, verified to really be a PDF, and posted as a multipart upload. One file per call. This is the operation for documents living behind authentication in a ticketing system, a document store or object storage.
Chat
2- Chat With PDF Asks a question about a previously added PDF and returns the answer generated from its contents. Earlier turns are passed in the Messages list because ChatPDF holds no state between calls, with at most six messages per request and older messages dropped server-side once the conversation grows past roughly two and a half thousand tokens. Include Page References adds bracketed page markers to the answer plus a structured list of the pages each claim was read from, and agents that write answers into a record of any consequence keep it on.
- Delete PDFs Permanently removes one or more previously added PDFs from the account, freeing the stored pages against the monthly allowance. Deleted sources can no longer be queried and the operation cannot be undone, so agents run it as a scheduled cleanup against sources the flow itself created rather than against a shared library.
Frequently Asked Questions
What can FlowRunner do with ChatPDF?
FlowRunner agents can run Add PDF From URL, Upload PDF File, and Chat With PDF in ChatPDF, plus 1 more action.
Does connecting ChatPDF to FlowRunner require OAuth?
No. ChatPDF connects to FlowRunner with an API key, no OAuth flow required.
Can ChatPDF trigger a FlowRunner workflow automatically?
ChatPDF doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with ChatPDF
$100 in credits. No card required. Connect in minutes.