Read the Text Printed in a Picture: the New image_ocr Endpoint
Opensolr could already tell you what a picture shows. As of today it can also tell you what is written on it, and that is the difference between a folder of photographed paperwork and a searchable archive of it.
01 What the two answers look like
Take a photo of a petrol receipt. Describing what the picture shows gets you as far as receipt, paper, document, text — true, and useless for finding that one receipt among four hundred. Reading what is printed on it gets you the station, the litres, the total and the fiscal number, which is what a person actually remembers.
receipt, paper, text, document
PETROM STATION 1042 · BENZINA STANDARD 95 · TOTAL 300.86 LEI · BON FISCAL 004417 · 16/09/2026
02 What it is good for
- Invoices and receipts. Index the supplier, the invoice number, the amount and the date off the picture, and an expense archive answers “that invoice from the plumber, about nine hundred” instead of needing a file name.
- Product and shelf labels. The article number and the product name come off the label, so a photo taken in a warehouse finds the product in a catalogue index.
- Screenshots. A screenshot becomes findable by the words in it, which is how most people remember screenshots.
- Business cards and documents. A name, a company and a phone number are searchable the moment the picture is indexed.
03 How to call it
One POST, up to five base64 pictures, one answer per picture in the order you sent them:
| Endpoint | POST https://opensolr.com/solr_manager/api/image_ocr |
|---|---|
| Sends | email, api_key, index_name, images (1–5, base64) |
| Answers | the text per picture, and the same reading split into lines |
| Costs | one AI request per picture actually read — a picture read before is answered from cache, for nothing |
Reading only pays off on pictures that carry text, so ask image_clip or image_index first and look at the new text_like field in the answer: it says whether the picture carries anything text-like, decided over the fifty best guesses the model considered. Send those pictures for reading and skip the rest.
04 Where it runs
The reading does not run on the GPU server that answers the embedding and language-model calls. It runs on a set of Opensolr Solr machines that do nothing else, which is why a bulk run of a few thousand pictures does not slow down anybody's searches or queue behind the AI work. Pictures are read in memory and deleted immediately; only the reading is kept, against the picture's own fingerprint, so the same picture is never read twice.