AI OCR at scale
2,800 scanned pages turned into clean text in about 25 minutes, page numbers kept for citation.
All work
Case study · Smart search · RAG
A whole library, searchable by meaning.
A specialist research library, mostly in Arabic and partly scanned. We digitised it, indexed it passage by passage and built an assistant that finds the right text by meaning — and always shows where each answer comes from.
- Sector
- Research & knowledge
- Our role
- OCR, search engine, assistant, hosting
- Status
- In production
- 40,000+passages searchable by meaning
- 8,000+pages from 7 sources
- 2,800scanned pages digitised by AI OCR
- 25 minto digitise all of them — none failed
The challenge
The answers were in the books. Finding them was the problem.
Thousands of pages, several sources and many scanned volumes that no search box could read. Researchers spent hours finding a passage they half remembered.
Keyword search fails on Arabic in particular: the same idea is written in many ways, and nobody knows the exact words of the passage they are looking for.
What we built
From scanned pages to answers with sources
Search by meaning
Every passage is indexed by meaning, so a question finds the right text even when it shares none of its words.
An assistant that cites
Answers are written from the passages found, each one with its source and page.
Arabic first
Designed for Arabic text from day one, with its spelling variants and classical style.
Measured, not guessed
AI models compared in blind tests on real questions before choosing the one in production.
On servers we control
The search index and the application run on our own infrastructure.
How it works
Four steps from a shelf of books to a cited answer
-
Digitise
Scanned volumes go through AI OCR; digital texts are cleaned.
-
Split
Texts are cut into passages, each tagged with its source and page.
-
Index
Each passage is turned into a vector and stored in a search index.
-
Answer
The assistant retrieves the best passages and answers with citations.
Results
Hours of reading, replaced by answers you can check
Client details are kept confidential. Full references and a live demo are available on request.
Built with
- Python
- FastAPI
- Qdrant
- Gemini
- OpenAI
- Claude
- nginx
- Answers in seconds instead of hours of reading
- Every answer traceable to a source and a page
- Scanned books finally searchable
- Questions understood however they are phrased
Want something like this?
Tell us what you have in mind. We'll suggest the simplest version that does the job, with a clear quote.
Start a project