Historical Records Extractor

Reads scans of old, handwritten and archival paper and gives you back clean text, one row per page.

This is built for the documents that ordinary text extraction gives up on. Census sheets with forty columns and a hundred cramped rows. Ship manifests written in a clerk's hurried cursive. Parish and church registers. Military service and casualty records. Land, probate and institutional ledgers. Faded, skewed, over-ruled, bled-through paper.

Every page comes back with a signed delivery receipt. The receipt binds the text we returned to the exact page image we were sent, so you can show later that a transcript is the one you were actually delivered, and check it without asking us anything.

You pay per delivered page, in USD. Pages that fail are not charged.

Who this is for

• Genealogists and family historians working through parish, census and immigration records
• Archivists and local history societies digitising a collection
• Records researchers who need a searchable transcript of a bound volume
• Legal, probate and land researchers reading old registers
• Academic historians building a dataset out of handwritten sources
• Anyone who has a folder of scans and needs the text out of them

If your documents are modern printed pages, a generic text extractor will be cheaper and will do fine. Come here for the difficult paper.

What you give it

Field	Required	Default	What it does
Document URLs	Yes	none	Direct links to the scans you want read. PDF, PNG, JPEG and WebP. One link per line.
Output format	No	Markdown	Markdown keeps reading order and turns tables into text tables. HTML keeps richer structure. JSON (layout blocks) returns the page as positioned blocks.
Max pages per document	No	50	Safety cap for big PDFs. Between 1 and 200. Pages past the cap are skipped and reported, never billed.
Include delivery receipts	No	On	Attaches the signed delivery record to every row.

Notes on input:

• The Actor reads files from a link. There is no upload box inside it. If your scans are on your computer, put them in a key-value store on this platform or on any file host first, then paste the direct links.
• PDFs are split into pages and each page is read on its own.
• A document that cannot be fetched, is larger than 100 MB, or is not one of the four accepted types is skipped. It is listed in the run summary with the reason. It is never silently dropped.

What you get back

One dataset row per page. A successful row carries:

Field	What it holds
source	The document link this page came from
page	Page number inside that document, starting at 1
status	ok
output_format	markdown, html or json, whichever you asked for
content	The extracted text for this page
billed	true if this page was charged, false if it was delivered without a charge
receipt	The signed delivery record for this page, present when receipts are switched on

A page that could not be read produces a row with status: failed, the source and page, and an error explaining what went wrong. Failed pages are not charged.

About billed: it is almost always true. It comes back false in one case, and we would rather tell you than hide it. If you set a maximum cost for the run and the run reaches it while a page is already being read, that page is finished and handed to you anyway, unbilled, because the work was already done. You keep the page and you are not charged for it.

The run also writes a SUMMARY record to the run's key-value store:

documents, pages_prepared, pages_settled, pages_settled_unbilled, pages_failed, pages_unprocessed_spend_limit, documents_skipped (with a reason for each), charged_pages, charging_disabled, output_format.

That summary is the honest count of what happened in the run. Use it to reconcile a big job against your invoice.

Signed delivery receipts

Each delivered page carries a small signed record. It contains a fingerprint of the page image we were sent, a fingerprint of the text we returned, how long the work took, an identifier for the service that produced it, a delivery reference, and a signature over all of it.

What that buys you:

• Proof of what was delivered. If a transcript is questioned months later, the receipt shows that this exact text was returned for that exact page image. Change one character of either and the fingerprints stop matching.
• A check that does not depend on us. The receipt travels inside your dataset. Verifying it is arithmetic on data you already hold, not a lookup against a service that has to still be running.
• A clean chain for a client or an archive. If you deliver transcripts to somebody else, you can hand over the receipts with them.

The receipt records the delivery. It is not a statement about accuracy, and it does not certify that the reading is correct. Nothing can do that except a human reading the page.

The receipt itself ships in every run today. If you want the step-by-step verification procedure and the signing key, ask through the Issues tab and we will send them.

Pricing

Pay per event, in USD.

• Per delivered page. You are charged once for each page that is read and returned with its receipt.
• Failed pages cost nothing. Skipped documents and pages past your page cap cost nothing.
• You can cap a run. Set a maximum total cost before you start it. When the run reaches the cap it stops taking new pages and shuts down cleanly, and you keep everything already delivered.

Start with two or three pages of your own material before you commit a whole collection. That is the only honest way to find out how this performs on your handwriting, your scanner and your century.

Limitations, stated plainly

We would rather you read this now than find it out during a run.

• Roughly one page per minute. Pages are processed one at a time today. A 50-page book is close to an hour of run time. Plan around it and use the page cap. We are working on it and this line gets updated when it changes, not before.
• Large documents are processed page by page. There is no whole-document mode. A table that runs across two pages comes back as two rows, and stitching it back together is your step.
• Each page is read on its own. Context does not carry from one page to the next, so a name spelled out on the cover page does not help disambiguate a scrawl on page nine.
• Scan quality decides the outcome. Faint ink, bleed-through from the reverse, tight bindings, skew and low-resolution phone photos all reduce what can be read. Higher-resolution scans read better.
• We publish no accuracy figure. We will not quote one for your material either, because we have not seen your material. Run a sample.
• No upload box. Files have to be reachable by a direct link.
• 100 MB per document. Anything bigger is skipped with a reason.

FAQ

Can it read old cursive handwriting? That is what it is built for. How well it does on any particular hand depends on the hand and the scan, so run a couple of pages before committing a collection.

What document types can I send? PDF, PNG, JPEG and WebP. PDFs get split into pages automatically.

How do I send files that are on my computer? Upload them to a key-value store on this platform, or to any file host, and paste the direct links. The Actor reads from links.

How long will my job take? Roughly a minute a page at the moment. Twenty pages is about twenty minutes. Very large jobs are better split into several runs.

What happens if a page fails? You get a row for it with status: failed and the reason. You are not charged for it. The rest of the run continues.

How do I stop a job from costing more than I want? Set a maximum total cost for the run. The run stops taking new pages once it reaches that, and shuts down cleanly rather than erroring.

Why is one of my rows marked billed: false? Because it was delivered after your run hit its cost cap, while it was already being read. You get the page for free. See the billed note above.

Do you keep my documents? Your files are fetched for the run, read page by page, and returned to your dataset. The delivery record carries fingerprints of the page and of the text, not the text itself.

Can you guarantee the transcript is correct? No, and neither can anyone else. Machine reading of handwritten historical paper produces a first pass that a human then corrects. What we guarantee is that the text you received is provably the text we delivered for that page.

What is the receipt actually for? Three things: proving to a third party what you were delivered, catching any later tampering with a transcript, and giving you a check you can run without us.

Is there a bulk or volume price? Ask through the Issues tab.

Data handling

Plainly, so there are no surprises:

• Your document content leaves the platform. Each page image is sent to an external extraction service, which performs the extraction and returns the text. That request carries the page image, its file type and the output format you chose. Nothing else from your input goes with it.
• A per-user identifier accompanies each request. Every request carries a one-way hash of the platform user id, plus the name of the venue the run came from, for operational accounting on the supply side. It is not reversible into your account name, email or any other profile field, but it is a stable pseudonymous identifier rather than anonymity: the same user produces the same value on every run, so requests from one user can be correlated with each other.
• Storage stays in your account. Extracted content and receipts are written to your own run's dataset and key-value store; this Actor keeps no copy of its own. What the external extraction service retains is governed by its own terms, not by this Actor.

About this service

This Actor is operated end to end by its publisher. Pricing is in USD.

Questions, problems and bug reports go through the Issues tab on this listing. That is the fastest route to us and it is the one we watch.