Which booking confirmation PDFs Kokai reads, and how

How it works: text, not image

A booking confirmation PDF almost always has real text inside: the kind you can select and copy. Kokai extracts it with pdf.js, Mozilla’s library (Apache-2.0 licence) that Firefox uses to display PDFs, and loads it only when you add a PDF. It all happens on the device:

1. If the file is over 20 MB, it’s stored without being read.
2. pdf.js extracts the text of the first 6 pages: confirmations put what matters at the top, and twenty pages of terms add nothing.
3. Kokai rebuilds the lines: pdf.js returns pieces of text with their position, and a line ends when the next piece drops by more than half a letter’s height.
4. The text is lowercased and stripped of accents to search for labels (“Número de confirmación” and “numero de confirmacion” are the same), but what’s found is cut from the original, so the booking reference keeps its capitals.
5. It decides what it is and who it’s from, and looks for the details by their labels. There are no per-provider templates, which change every few months.
6. If the class isn’t recognised and fewer than two details come out, it counts as not read: «No he encontrado los datos de una reserva en el PDF. Rellénalos a mano.» (“I couldn’t find booking details in the PDF. Fill them in by hand.”)

There’s no text recognition in images (OCR). A photo of the ticket, a screenshot or a scanned PDF is stored and opens offline, but isn’t read. Nor is a password-protected PDF.

Where the app reads a PDF

In Documents, while reading, the app says «Leyendo el PDF en este dispositivo…» (“Reading the PDF on this device…”) (in Accommodation, «Leyendo el PDF…»). What it understood appears in a line starting with «Leído en este dispositivo:» (“Read on this device:”). More about documents in Offline travel documents.

The 8 booking classes

Each class sets the document type, the tag that gets added and which labels are searched for the dates. To recognise it, typical words count, each one only once even if it repeats:

The 65 brands it recognises

A recognised brand adds its tag to the document («Booking.com», «Renfe») and weighs heavily when deciding the class. If yours isn’t listed, the booking can still be read by its words and labels: only the provider tag is missing.

Some brands only count in their full form, because their bare name is an everyday word: “Booking.com” or “Booking B.V.” (a booking is just a reservation), “Italo Treno”, “Enterprise Rent-A-Car”, “Avis Budget” or “alsa.es”. The bare name (“booking”, “alsa”, “avis”) is a weak mention.

How it decides what it is and who it’s from

The class and the provider come from a points tally, the same for every PDF:

• A class’s score is its words plus its best brand.
• The class with the most points wins, and only if it has at least 4: two typical words are enough (a hotel with no platform is recognised as accommodation), a lone “hotel” isn’t.
• The provider is named if its brand, in that class, has at least 5 points: its full name has to appear, or be in the file name. Three weak mentions add up to 3 and aren’t enough.

Which details it extracts

The title is built according to the class: the hotel’s name (or «Alojamiento de Booking.com», “Booking.com accommodation”, if it doesn’t appear), «Vuelo» with its number, or the class and the provider («Tren · Renfe»).

The labels it recognises, language by language

They’re matched ignoring case and accents. A dash means that language has no form of its own in the list (it often matches another: the Catalan “número de reserva” is the Spanish form).

As a last resort, the price is also searched for after the bare word in any of the seven languages: precio, price, importe, amount, montant, prix, prezzo, preço, preu.

Dates and times

• Abbreviations that are also words (out, des, set, gen, mag, lug, ott, mai, mar) only count as a month if a year follows: “3 out of 5 guests” isn’t October 3 (“out” is October in Portuguese).
• In a table with “Check-in Check-out” on one line and both dates below, the first date after “Check-out” is the check-in date, and it’s skipped.
• If there’s no label, it uses the first date that doesn’t follow “fecha de reserva”, “emitido”, “booked on”, “fecha de nacimiento” or “caduca” (booking date, issued, date of birth, expires). For insurance, it doesn’t guess.
• Accommodation with “3 noches” (3 nights) and no readable check-out date: check-out is worked out by adding the nights to check-in.
• An end date more than 62 days after the start is dropped: it’s some other date that slipped in.
• Times: 15:00, 15h00, 3:00 PM or 3 PM. For the start time, the one with “desde”, “from” or “a partir de” wins; for an accommodation’s check-out, the last one with “hasta”, “until” or “before”: “de 07:00 a 11:00” (from 07:00 to 11:00) means check-out until 11:00.

Price and currency

The currency comes from the symbol or code next to the amount, before or after it (“€ 390,00”, “390,00 EUR”), or from the code in brackets after “Total (EUR)”. It recognises 37 ISO 4217 codes:

EUR USD GBP JPY CHF CAD AUD MXN ARS COP CLP PEN BRL DKK SEK NOK PLN CZK HUF
RON BGN TRY MAD THB CNY HKD SGD NZD ZAR INR KRW ISK IDR VND EGP AED ILS

Separators are read as each country writes them: 1.234,56 and 1,234.56 are the same, and so is 1 234,56; “1.234” is one thousand two hundred and thirty-four, and “12,50” is twelve point five.

A bare “$” is always read as US dollars and a “¥” as yen. If you book in Mexican pesos or Chinese yuan and the PDF only shows the symbol, correct the currency on the stay. If there’s no currency, the stay takes the trip’s.

What it does with what it reads

There’s one rule: it never overwrites what someone has typed.

• On a new document, if you change the type or title while it’s reading, yours stays. If you change the type that was read (a «Reserva», “Booking”, that you turn into «Entrada», “Entry ticket”), the tags from the reading aren’t added.
• On an existing document, only the booking reference, holder and dates are filled in if they’re empty, and the type only if it was «Otro» (“Other”). If the type read is different, the card offers «Cambiar el tipo a…» (“Change the type to…”) without changing it.
• Two tags are added, the provider’s and the class’s («Booking.com» and «Alojamiento»), without removing the ones it had.

If it’s accommodation with check-in and check-out, and the document is a booking, it’s tied to a stay, that hotel’s nights on the trip:

1. If there’s already a stay linked to that document, it’s completed (times, price, cancellation) without touching what it has.
2. If not, it looks for the hotel: the document’s related place, or an accommodation on the map with the same name (at least 60% of the words in common, not counting “hotel”, “apartamento”, “de” or “the”).
3. If that hotel has an option with no linked document that matches those nights, it becomes booked, with the document linked and the dates from the PDF.
4. If the hotel isn’t on the map, its address goes to Nominatim (OpenStreetMap’s search engine), or its name within the destination area. If you already have accommodation less than 150 m from the result, it’s that one; if not, the place is created with its address and phone, and its photos are looked up.
5. If there was no option, a new booked stay is created, at that place if it was found. Offline, or if Nominatim finds nothing, it stays unplaced, with its name, and you place it by hand later.

The stay brings the accommodation reminders: free cancellation, «Mañana duermes en…» (“Tomorrow you sleep at…”) and «Hoy dejas…» (“Today you leave…”) (full table in Documents). If a booking document has dates but wasn’t tied to a stay, its card offers «Es un alojamiento: añadir estas noches» (“It’s accommodation: add these nights”).

Nominatim only receives the address text (or the hotel name and the area), with at least 1.1 seconds between requests, as its usage policy asks. Never the booking reference, the holder or the price.

Example: a Booking.com confirmation for Lisbon

Limitations

• No photos or scans. Only text is read; there’s no OCR.
• Tested with sample texts, not with real confirmations from each provider. The tests mimic what pdf.js extracts from Booking.com (in Spanish and English, with a table), Airbnb (in both languages), a hotel with no platform, Ryanair, Renfe and Europcar, and a generated PDF is read from start to finish. It hasn’t been tested with each of the 65 brands.
• A PDF with jumbled text (columns that pdf.js returns in a different order) can leave labels far from their values, and the detail doesn’t come out.
• Only one ticket per document. A return-trip PDF gives the outbound and return dates, but not two documents, nor stopovers.
• Only the first PDF when you choose several files at once while creating a document.
• Everything read can be corrected. Always look at the card: the app asks you to with «Revisa que esté bien.»

If your booking wasn’t read properly

Correct the details by hand on the card: they save themselves and the reading won’t overwrite them again. And if you’d like the next version to read it better, tell us in the contact form with the reason “A PDF booking that wasn’t read properly”. It helps to say:

• The provider and the language of the confirmation.
• Which detail failed (the check-out date, the price, the hotel name) and what the PDF said next to it, with the label exactly as written.
• Whether the PDF is a printed email, a download from the provider’s website or a boarding pass.

Don’t paste real booking references, names, addresses or card numbers: swap them for made-up ones before sending (“XXXX1234”, “Ana Ejemplo”). The form doesn’t accept attachments, and the PDF isn’t needed: the lines that fail are enough.

Frequently asked questions

Sources and references

The original documentation for what is cited here, checked on this page’s review date.

1. PDF.js Mozilla
2. PDF.js repository (code and Apache-2.0 licence) Mozilla, on GitHub
3. ISO 4217, currency codes (maintenance agency) SIX Group
4. Nominatim Usage Policy OpenStreetMap Foundation
5. Nominatim: Search API Nominatim

Keep reading

• DocumentsHow Kokai keeps tickets, bookings and passport as PDF or photo to open offline, what it warns you about before they expire and what stays on your device.
• AccommodationHow Kokai organises accommodation by night: bookings, undecided options, gaps, double bookings, free cancellation alerts and the hotel address offline.
• Search for accommodationHow to build Booking.com, Airbnb, Expedia, Vrbo, Agoda, Trip.com, Kayak and Google Hotels search URLs with destination, dates and guests, with examples.
• GlossaryTravel glossary: what a geofence, a stay, a booking reference, GPS hysteresis or PMTiles is. 35 travel and map terms, with the figures Kokai uses.