PDF Revision Monitor - Text, Pages & Change Evidence

See the evidence in a PDF revision

For teams checking recurring reports and documents: extract the text layer of public PDFs with page references and hashes, then optionally compare complete observations. Bounded download and extraction controls keep the scope explicit.

Open this Actor in Console · First run · Results · Cost · Full guide

Try this Actor

• First run: Run the existing Quick start example, {}, to inspect the single Census retail PDF supplied by the default. Review its extraction completeness before enabling comparison.
• Read the results: Open the default Dataset for document observations and SUMMARY for processing status. Check Complete extraction, the extracted-page count and the baseline status before relying on a comparison.
• Keep in mind: There is no OCR, visual layout analysis or table-cell extraction. A byte-only change may be formatting or metadata; partial extraction does not replace a complete comparison baseline.

Extract Current Census Retail Report — copy this input into the JSON editor in Console.

{}

Result preview

The dataset view highlights these columns:

• PDF URL · url
• Document ID · documentId
• Total pages · pageCount
• Pages extracted · extractedPageCount
• Characters extracted · characterCount
• Complete extraction · completeExtraction

Use the original source links and the field definitions below when interpreting the results. CSV works well for flat columns; JSON preserves nested values.

Usage and pricing

Base result rate on the FREE tier: $10 per 1,000 document observations. The Pricing tab shows your plan's result rate and any Actor Start charge. Set a small result limit for the first run; the requested limit is not a guarantee of that many results.

---

Full guide

Extract text from public PDF URLs with page citation links and stable content hashes. Optionally compare each document with its previous complete observation to identify changed pages, removed pages or byte-only changes. No source API key or OCR service is required.

Quick start and exact examples

Default input extracts the current public Census retail-report PDF:

{}

Keep your own report identity and external reference:

{"documents":[{"url":"https://www.census.gov/retail/marts/www/marts_current.pdf","documentId":"retail-report","externalId":"research-retail-001"}],"maxPagesPerDocument":100}

Compare future observations at a stable report URL:

{"documents":[{"url":"https://www.census.gov/retail/marts/www/marts_current.pdf","documentId":"census-monthly-retail-release"}],"comparePrevious":true,"baselineKey":"example-census-report"}

The first complete comparison run creates a baseline. Its row says comparisonStatus: no_previous_baseline; SUMMARY.documents[].baselineUpdated confirms the actual history commit. A partial first observation cannot create history and says changeType: partial_observation. Unchanged successful checks still return and bill an observation. Source publication and revision statements remain in the extracted text; a URL named “current” does not establish that every figure is the latest revised estimate.

Input and limits

documents accepts up to 20 public URL strings or objects with url, optional documentId, and optional externalId. The default identity is the supplied URL. Supply the same intentional documentId to compare versions whose download URL changes. IDs must be unique within a run.

• maxPagesPerDocument: 100 by default and at most 100; pages are read starting at page 1. A shorter cap produces a partial observation.
• maxFileSizeMb: 10 MiB by default, at most 20. Larger responses are rejected before parsing.
• maxCharactersPerDocument: 250,000 by default, at most 1,000,000. Reaching the character limit makes extraction partial.
• maxRunSeconds: 240 by default, also bounded by the platform deadline minus 25 seconds. A document parser is limited to at most 60 seconds and the actual remaining budget.

Only public HTTP(S) ports 80/443 are allowed. DNS results and redirect destinations are checked; robots restrictions are honored. Downloads need no custom headers, cookies, passwords or credentials. Parsing runs in a bounded worker without evaluating embedded PDF scripts.

Output

One dataset row represents one PDF with usable extracted text:

• url, inputUrl, documentId, externalId, observedAt: source identity and observation time.
• byteLength, byteHash, contentHash: exact downloaded bytes and normalized page-text identity.
• pageCount, extractedPageCount, characterCount, completeExtraction, limitsReached, noTextPages: actual coverage.
• pages: page number, extracted text, text hash and citationUrl ending in #page=N.
• metadata: selected source PDF metadata when present. These values are not verified publication dates.
• comparisonStatus, previousObservedAt, changeType, changedPages, removedPages, unchangedPageCount: comparison evidence. Before/after excerpts are bounded previews of the page, not an exact inline word diff.

changeType distinguishes text_changed, bytes_only_changed, unchanged, baseline, partial_observation and not_compared. Page comparisons align by page number: inserting a page can shift later pages and create several changed-page entries. Byte-only changes can reflect metadata or formatting; no visual-change analysis is claimed.

History and failure behavior

comparePrevious is opt-in. History is stored in maydit-pdf-revision-baselines-v1, keyed by document ID, namespace and extractor version. A complete baseline stores hashes, page numbers and bounded text previews. It never stores a second copy of the PDF bytes.

Only a completely parsed document whose output was saved advances history. Partial/capped/no-text pages and failed output preserve the earlier complete baseline; removals are never inferred from a partial extraction. A saved final row that reaches the spending cap still counts and can advance complete history. Use one writer per document ID and namespace: the pre-write conflict check is not an atomic lock.

Partial documents with some usable text return clearly marked observations. All-textless, encrypted, malformed, blocked, or oversized documents are failures reported in SUMMARY, with no fake dataset row. A run with no usable text fails; a normal spending-limit stop succeeds with status: budget_limited, including when no row was affordable.

Supported content and pricing

This reads a PDF's text layer. It does not perform OCR, handwriting recognition, image extraction, semantic legal review, or structured table-cell extraction. Reading order in multicolumn layouts can be imperfect. Empty page text does not prove that a page is visually blank. Text hashes describe the extracted representation, not the visual document.

The billing unit is one saved document observation, including unchanged and explicitly partial observations. Individual pages do not add billable rows. Consult the Store Pricing tab for current prices; failed documents create no result charge. Platform/start charges, if listed, are separate.

Sources are user-supplied public PDFs. Use documents you are authorized to process. Parsing uses Mozilla PDF.js; the default is a real U.S. Census report.

Help with your run

Open the Issues tab on this Actor with the run URL, the input field that caused trouble and what you expected to receive. Remove API keys and private information before posting.

Explore more Mayd-It Actors

Understand your run

Open Run report in the run’s Output tab for saved-record counts, reported source issues and limits, and suggested next steps. Use CSV export to take the saved results into a spreadsheet. The existing JSON results and detailed machine-readable summary remain available.

The report reflects this run’s recorded coverage. A successful run or a saved result does not establish that every possible source record was collected.