PDF Vector
Document ProcessingParse documents, answer questions about them, and extract structured data with PDF Vector, converting PDFs, Office files, and images into clean Markdown. Agents read a document and act on what it says.
What This Integration Enables
PDF Vector is organised around a grid rather than a list, and understanding the grid is what makes it useful. There are three verbs. Parse converts a document into clean Markdown. Ask answers a natural-language question about it. Extract returns JSON matching a JSON Schema you supply, guided by an extraction prompt. Then there are four document families, each with its own tuned model: general documents, invoices, identity documents and bank statements. Every verb is available for every family, so the choice an agent makes is not "which endpoint" but "which question am I asking, about which kind of document".
That matters because the tuned models are not a marketing distinction. Parse Invoice reads line items, totals and tax fields differently from the generic parser. Parse Identity Document handles machine-readable zones and returns the document type it detected. Parse Bank Statement preserves transaction tables and running balances that a general parser flattens. Reaching for Parse Document on an invoice is a decision with a cost, and an agent that knows the difference is worth more than one that treats every PDF the same. Everything runs synchronously and reports the credits each call consumed, which gives a flow a real handle on spend without a separate accounting step. There is also a research suite covering literature search, citation graphs and grants, which sits outside the document pipeline but shares the same key.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Onboarding a borrower from the documents they uploaded
An applicant submits a form with a passport and three months of statements. The agent runs Extract Identity Document Data against the passport with a schema naming exactly the fields the case file wants: full name, document number, date of birth, nationality, expiry. It runs Extract Bank Statement Data across the statements for the account number, statement period, opening and closing balances and the transaction array. It then compares the extracted name and date of birth against what the applicant typed on the form. Where everything agrees and the ID has real time left on it, the case file in Salesforce is populated and the application moves to the next stage. Where anything disagrees, it does not.
Accounts payable that reads the right way
Supplier invoices arrive in a shared Dropbox folder. Rather than parsing them generically, the agent runs Extract Invoice Data with a schema matching the ledger's bill object, because the invoice-tuned model is measurably better on the parts that matter here: line items, tax and totals. Where a document raises a question the fixed schema does not answer, the agent follows up with Ask Invoice, which handles things like whether a purchase order number is present or what the stated payment terms are, without a second schema being written. The bill posts to QuickBooks Online with the source document attached.
Turning a document library into something an agent can read
A policy archive of Word files, scanned PDFs and PowerPoint decks needs to become searchable text for a downstream retrieval step. The agent runs Parse Document over each file, which returns clean Markdown and a page count and handles tables, handwriting, math and multilingual content at the higher model tiers. The Markdown is written into Confluence and indexed. Because each call reports the credits it consumed, the agent tallies the cost of the backfill as it runs and reports the total rather than discovering it on an invoice.
Human-in-Loop Highlight
An identity document is not a document. It is a person, and every field on it is a claim about that person that something downstream will act on. Extract Identity Document Data returns a name, a document number, a date of birth, a nationality and an expiry date, all typed and all clean-looking, and the two ways it can be wrong pull in opposite directions and are both bad. An expiry date read with one digit off either admits an expired document into an onboarding decision or blocks a real customer whose passport is perfectly valid. A name transliterated slightly differently from the application form is either a genuine mismatch worth stopping on or an accent mark the model normalised away.
Neither of those is a call an agent should close. So the agent does everything up to the decision and then stops. It runs the extraction, compares every field against the application record, tests the expiry against the policy window, and where anything disagrees it puts the comparison in front of a compliance reviewer with the document image beside it: "The passport reads SOARES DA SILVA, MARIANA with expiry 2026-11-03. The application says Mariana Soares. The name is a partial match and the expiry falls inside the ninety-day window. Confirm the applicant and the document, request a renewed ID, or reject?" The reviewer decides, and the decision attaches to the case file with the extraction and the image that produced it. The human-in-the-loop step here is not a delay in the process. It is the only part of the process that carries the liability, and the agent's job is to hand it over already prepared.
Agent Capabilities
20 actionsDocuments
3- Parse Document Converts a document into clean Markdown and returns its page count. Supports PDF, Word, Excel, PowerPoint, CSV, images, plain text, RTF, HTML, OpenDocument, EPUB and bibliography files, supplied as a public URL or a FlowRunner file. Higher model tiers handle tables, handwriting, math and multilingual content, and the top tier also returns an HTML representation. Reports the credits consumed.
- Ask Document Parses a document and answers a natural-language question about it, returning the answer as Markdown. Accepts large files via a public URL or a FlowRunner file. Higher model tiers suit nuanced questions requiring deeper reasoning over complex documents.
- Extract Document Data Extracts structured JSON from a document according to a JSON Schema you supply, guided by an extraction prompt, returning an object matching that schema exactly. The route for turning contracts, reports and forms into typed records the receiving system already understands.
Invoices
3- Parse Invoice Parses an invoice with an invoice-tuned model, returning Markdown plus the page count. Accuracy on line items, totals and tax fields is higher than the generic document parser, so this is the right call whenever the document is known to be an invoice. The top tier additionally returns full HTML preserving the original layout.
- Ask Invoice Answers a natural-language question about an invoice using the invoice-tuned model. Suits the questions a fixed schema does not cover: the amount due, the payment terms, whether a purchase order number appears anywhere on the document.
- Extract Invoice Data Extracts structured invoice data matching a JSON Schema you supply, using the invoice-tuned model. Built for accounts payable, where invoice numbers, vendor details, line items, tax and totals must land in fixed fields rather than in prose.
Identity Documents
3- Parse Identity Document Parses a passport, driver's licence or national ID card, returning Markdown along with the detected document type. Uses an identity-tuned model for better accuracy on machine-readable zones and ID fields.
- Ask Identity Document Answers a natural-language question about an identity document, such as the expiry date, the nationality, or whether the document is still valid. Returns the answer as Markdown.
- Extract Identity Document Data Extracts structured identity data matching a JSON Schema you supply: full name, document number, date of birth, nationality, expiry date and whatever else the schema names. Designed for onboarding flows that need consistent typed fields, and the operation whose output an agent should never act on without a reviewer where anything disagrees with the application record.
Bank Statements
3- Parse Bank Statement Parses a bank statement with a statement-tuned model, returning Markdown plus the page count. Preserves transaction tables, running balances and account headers more reliably than the generic parser.
- Ask Bank Statement Answers a natural-language question about a statement, such as the closing balance, the total spend in a category, or the largest transaction in the period.
- Extract Bank Statement Data Extracts structured statement data matching a JSON Schema you supply: account number, statement period, opening and closing balances, and an array of transactions. The operation reconciliation and bookkeeping automations are built on.
Academic Research
7- Search Papers Searches academic publications across multiple scholarly databases with unified relevance ranking, covering Semantic Scholar, PubMed, arXiv, Google Scholar, ERIC, Europe PMC, OpenAlex and Crossref. Returns titles, authors, abstracts, DOIs, citation counts and PDF URLs, with optional publication-year filtering.
- Fetch Papers Fetches specific publications by identifier with automatic provider detection and batch processing, accepting DOIs, PubMed IDs, arXiv IDs and several others in a single call. Returns full metadata plus a per-identifier errors list for anything that could not be resolved.
- Parse Paper Resolves a paper from an identifier or a publisher URL to a publicly available PDF, then parses it into Markdown using the same model tiers as Parse Document. Returns a not-found result when the paper cannot be located or no public PDF exists, which agents branch on rather than treating as an error.
- Find Citations Finds supporting citations for a paragraph of text by splitting it into sentences and scoring candidate papers against each one with semantic similarity, returning ranked citations with a relevance score. Useful for adding references to draft writing.
- Get Paper Graph Retrieves the citation graph around a paper: the papers citing it and the papers it references, each as a full publication object. Both directions support independent limits and pagination.
- Find Similar Papers Finds papers similar to a seed paper using citation network analysis rather than keyword matching, ranked by a combined similarity score that includes citation-network importance and graph distance. Optionally returns edge data for building a visualisation.
- Search Grants Searches research grants and funding opportunities across Grants.gov, NIH RePORTER, CORDIS and UKRI with unified results and relevance ranking. Filters by application deadline window and funding amount, returning agency, programme, eligibility, funding range, deadlines and principal investigator details.
Account
1- Validate API Key Verifies the configured key is accepted by PDF Vector, returning success or an error describing the rejection. Used as a connection health check before running credit-consuming operations. PDF Vector exposes no credit-balance endpoint, so the credits consumed by each call are reported in that call's own response and an agent that wants a running total keeps it itself.
Frequently Asked Questions
What can FlowRunner do with PDF Vector?
FlowRunner agents can run Parse Document, Ask Document, and Extract Document Data in PDF Vector, plus 17 more actions.
Does connecting PDF Vector to FlowRunner require OAuth?
No. PDF Vector connects to FlowRunner with an API key, no OAuth flow required.
Can PDF Vector trigger a FlowRunner workflow automatically?
PDF Vector doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with PDF Vector
$100 in credits. No card required. Connect in minutes.