The itinerary data extraction engine extracts travel-related information from input in various forms, from PDF documents to ticket barcodes, from emails to calendar events, and provides that in a machine-readable way.
For linked class names read this in the API docs.
Various QML-compatible value classes based on that can be found in the
src/lib/datatypes sub-directory. Those do not implement the schema.org ontology one to one though, but focus on a subset relevant for the current consumers. Any avoidable complexity of the ontology is omitted, which mainly shows in a significantly flattened inheritance hierarchy, and stricter property types. This is done to make data processing and display easier.
There is one notable extension to the schema.org model, all date/time values support explicit IANA timezone identifiers, something that JSON cannot model out of the box.
De/serialization is provided via KItinerary::JsonLdDocument.
Input data is transformed into a tree of document nodes (KItinerary::ExtractorDocumentNode). This allows handling of arbitrarily nested data, such as an email with a PDF attached to it which contains an image that contains a barcode with an UIC 918.3 ticket container, without extractors having to consider all possible combinations.
A document node consists of a MIME type and its corresponding data, and potentially a number of child nodes.
Data extraction is then performed on that document tree starting at the leaf nodes, with results propagating upwards towards the root node.
- PDF documents, represented as KItinerary::PdfDocument.
- Emails, represented as KMime::Message.
- Apple Wallet passes, represented as KPkPass::Pass.
- iCal calendars and iCal calendar event, represented as KCalendarCore::Calendar and KCalendarCore::Event.
- HTML and XML documents, represented as KItinerary::HtmlDocument.
- UIC 918.3/918.9 ticket barcodes, represented as KItinerary::Uic9183Parser.
- European Railway Agency (ERA) FCB ticket barcodes, represented as KItinerary::Fcb::UicRailTicketData.
- European Railway Agency (ERA) SSB ticket barcodes, represented as KItinerary::SSBv1Ticket, KItinerary::SSBv2Ticket and KItinerary::SSBv3Ticket.
- IATA boarding pass barcodes, represented as KItinerary::IataBcbp.
- VDV eTicket barcodes, represented as KItinerary::VdvTicket.
These are primarily needed for internal use.
- Images, represented as QImage.
- Apple property lists (plist), represented as KItinerary::PListReader.
- HTTP responses, represented as KItineary::HTTPResponse.
These capture everything not handled above.
- JSON, represented as QJsonArray.
- Plain textual data, represented as a QString.
- Arbitrary binary data, represented as a QByteArray.
Data extraction is performed on the document tree starting at the leaf nodes, with results propagating upwards towards the root node. This means that results from child nodes are available to the extraction process, and can be extended/augmented there for example.
The entry point for data extraction is KItinerary::ExtractorEngine.
There's a number of built-in generic extractors for the following cases:
- The various ticket barcode types (IATA, UIC 918.3/9, ERA FCB, ERA SSB).
- Structured data in JSON-LD or XML microdata format included in HTML documents or iCal events.
- PDF flight boarding passes.
- Apple Wallet passes for flights, trains or events.
- iCal calendar events (depends on KItinerary::ExtractorEngine::ExtractGenericIcalEvents).
- ActivityPub events and places.
To cover anything not handled by this, there are vendor-specific extractor scripts. Those can produce complete results or merely fix or augment what the generic extraction has produced.
Extractor scripts consist of two basic parts, the filter defining when it should be triggered and the script itself (see KItinerary::ScriptExtractor). This is necessary as running all extractor scripts against a given input data would be too expensive. Filters therefore don't need to be perfect (noticing in the script it triggered on the wrong document is fine), but rather fast.
A number of additional processing steps are applied to extracted data (see KItineary::ExtractorPostProcessor).
- Simplify whitespaces in human-readable strings.
- Separate postal codes in addresses.
- Remove name prefixes.
- Convert human-readable country names into ISO 3166-1 alpha 2 country codes.
- Apply timezones to date/time values.
- Identify IATA airport codes based on airport names.
- Geographic coordinates based on IATA airport codes as well as a number of train station code.
- Timezones based on geographic coordinates, or where sufficiently unique country/region information.
- Countries and regions based on geographic coordinates.
- Countries based on international phone numbers (needs libphonenumbers).
If the result set contains multiple elements, merging elements referring to the same incidence is attempted. Two cases are considered:
- Elements that are considered to refer to exactly the same incidence are folded into one.
- An element referring to a location change from A to B and two elements referring to a location change from A to C and C to B are considered to refer to the same trip, with the first one providing a lower level of detail. The first element is folded into the other two in that case.
In the final step all results are checked for containing a bare minimum of information (e.g. time and name for an event), and for being self-consistent (e.g. start time before end time). Invalid results are discarded. See KItinerary::ExtractorValidator.
Extractor scripts are searched for in two locations:
- In the file system at
- Compiled into the binary at
Those locations are searched for JSON files containing one or more extractor script declarations.
The above example shows a single script declarations, for declaring multiple scripts in one file this can also be a JSON array of such objects. The individual fields are documented below.
Extractor filters are evaluated against document nodes. This can be the node the extractor script wants to process, but also a descendant or ancestor node.
An extractor script filter consists of the following four properties:
mimeType: the type of the node to match
field: the property of the node content to match. This is ignored for nodes containing basic types such as plain text or binary data.
match: a regular expression
scope: this defines the relation to the node the script should be run on (Current, Parent, Children, Ancestors or Descendants).
Anything attached to an email sent by "booki[email protected]". The field matched against here is the
From header of the MIME message.
Documents containing a barcode of the format "FNNNNNNNN". Note that the scope here is
Descendants rather than
Children as the direct child nodes tend to be the images containing the barcode.
PDF documents containing the string "My Ferry Booking" anywhere. This should be used as a last resort only, as matching against the full PDF document content can be expensive. An imprecise trigger on a barcode is preferable to this.
Apple Wallet passes issued by "org.kde.travelAgency".
iCal events with an organizer email address of the "kde.org" domain. Note that the field here accesses a property of a property. This works at arbitrary depth, as long as the corresponding types are introspectable by Qt.
A (PDF) document containing an IATA boarding pass barcode of the airline "AB". Triggering vendor-specific UIC or ERA railway tickets can be done very similarly, matching on the corresponding carrier ids.
A node that has already existing results containing a reservation from "My Transport Operator". This is useful for scripts that want to augment or fix schema.org annotation already provided by the source. Note that the mimeType "application/ld+json" is special here as it doesn't only trigger on the document node content itself, but also matches against the result of nodes of any type.
Extractor scripts are defined by the following properties:
script: The name of the script file.
function: The name of the JS function that is called as the entry point into the script.
mimeType: The MIME type the script can handle.
filter: A list of extractor filters as described above.
Extractor scripts are run against a document node if all of the following conditions are met:
mimeTypeof the script matches that of the node.
- At least one of the extractor
filterof the script match the node.
The script entry point is called with three arguments (this being JS, some of those can be omitted by the script and are then silently ignored):
- The first argument is the content of the node that is processed. The data type of that argument depends on the node type as described in the document model section above. This is usually what extractor script are most concerned with.
- The second argument is the document node being processed (see KItinerary::ExtractorDocumentNode). This can be useful to access already extracted results on a node (e.g. coming from generic extraction) in order to augment those.
- The third argument is the document node that matched the filter. This can be the same as the second argument (for filters with
scope= Current), but it doesn't have to be. This is most useful when triggering on descendant nodes such as barcodes, the content of which will then be incorporated into the extraction result by the script.
The script entry point function is expected to return one of the following:
- A JS object following the schema.org ontology with a single extraction result.
- A JS array containing one or more such objects.
- Anything else (including empty arrays and script errors) are considered an empty result.
API for supporting schema.org output:
- KItinerary::JsApi::JsonLd: factory functions for schema.org objects, date/time parsing, etc
API for handling specific types of input data:
- KItinerary::JsApi::ByteArray: functions for dealing with byte-aligned binary data, including decompression, Base64 decoding, Protcol Buffer decoding, etc.
- KItinerary::JsApi::BitArray: functions for dealing with non byte-aligned binary data, such as reading numerical data at arbitrary bit offsets.
- KItinerary::JsApi::Barcode: functions for manual barcode decoding. This should be rarely needed nowadays, with the extractor engine doing this automatically and creating corresponding document nodes.
API for interacting with the extractor engine itself:
- KItinerary::JsApi::ExtractorEngine: this allows to recursively perform extraction. This can be useful for elements that need custom decoding in an extractor script first, but that contain otherwise generally supported data formats. Standard barcodes encoded in URL arguments are such an example.
KItinerary Workbench allows interactive development of extractor scripts.
Let's assume we want to create an extractor script for a railway ticket which comes with a simple tabular layout for a single leg per page, and contains a QR code with a 10 digit number for each leg.
As a filter we'd use something similar as example 2 above, triggering on the barcode content.
The above example produces and entirely new result. Another common case are scripts that merely augment an existing result. Let's assume an Apple Wallet pass for a flight, the automatically extracted result is correct but misses the boarding group. The filter for this would be similar to example 4 above, triggering on the pass issuer.
A large number of real-world examples can also be found in the
src/lib/scripts folder of the source code or browsed here.
Using the C++ API is the most flexible and efficient way to use this. This consists of three steps:
- Extraction: This will attempt to find relevant information in the given input documents, its output however can still contain duplicate or invalid results. There are some options to customize this step, e.g. trading more expensive image processing against finding more results, depending on how certain you are the input data is going to contain such data. See KItinerary::ExtractorEngine.
- Post-processing: This step merges duplicate or split results, but its output can still contain invalid elements. The main way to customize this step is in what you feed into it. For best results this should be all extractor results that can possibly contain information for a specific incident. See KItinerary::ExtractorPostprocessor.
- Validation: This will remove and remaining incomplete or invalid results, or results of undesired types. For this step you typically want to set the set of types your application can handle. Letting incomplete results pass can be useful if you do have an existing set of data you want to apply those too. See KItineary::ExtractorValidator.
In cases where integrating with the C++ API isn't possible or desirable, there's also a command line interface to this,
This reads input data from stdin and outputs schema.org JSON with the results.
For easier deployment, the command line extractor can also be built entirely statically. This is available directly from the Gitlab CI/CD pipeline on demand. Nightly Flatpak builds are also available from KDE's nightly Flatpak repository:
Contribution of new extractor scripts as well as improvements to the extractor engine are very welcome, preferably as merge request for this repository.
Another way to contribute is by donating sample data. Unlike similar proprietary solutions our data extraction runs entirely on your device, so we never get to see user documents and thus rely on donated material to test and improve the extractor.
Samples can be sent to email@example.com@firstname.lastname@example.org@m.g and will not be published. Anything vaguely looking like a train, bus, boat, flight, rental car, hotel, event or restaurant bookings/tickets/confirmations/cancellation/etc is relevant, even when they are seemingly already extracted correctly (in many cases there are non-obvious details we don't cover yet correctly). If possible, please provide material in its original unaltered form, for emails the easiest way is "Forward As Attachment", inline forwarding can destroy relevant details.
Feel free to join us in the KDE Itinerary Matrix channel!