Skip to content

Catalogue
Where to start
How it works
Contribute
About

Language : Français

Where to start

Eight questions, two minutes, to find out whether your task needs AI and which kind. Nothing is sent anywhere: it is all computed in your browser, and no model is called.

What you want to do
Broadly, what do you want to do?
The verb, not the solution. If two answers fit, take the one that describes the result you expect.
- Detect and filterHow do I catch what should not get through?
- ExtractHow do I get that value out of the text?
- Classify and routeWhere should this message go?
- SearchHow do I find the right thing in all this?
- RecommendWhat do I show next?
- PredictWhat is going to happen, and what looks wrong?
- GenerateHow do I produce this without writing it by hand?
- TransformHow do I put this in another form?
- Recognise and transcribeHow do I read what is not text yet?
- Decide and validateDo I accept this, and on what grounds?
- I don’t know yet
How often does this task happen?
This is what decides the cost: a rare task and a continuous one are not solved the same way.
- A few times a day, or less
- Hundreds of times a day
- Thousands of times a day, or continuously
- I don’t know
Your task
Which of these tasks is closest to yours?
These are the tasks the catalogue already covers in this family.
- Offer suggestions from the first characters typed into a search bar.Autocomplete a search bar
- Check that a SIREN or a SIRET typed into a form is well-formed, before storing it.Check a company identification number
- Know what an uploaded file really holds, before accepting it, storing it or serving it back.Check a file is really the format it claims
- Decide whether a password chosen at sign-up can be accepted, without letting it leave your infrastructure.Check a password against a policy
- Check that an IBAN, typed or pasted, is well-formed, before the transfer goes out.Check bank details before a transfer
- List the addresses quoted in a web page, resolved to absolute ones.Collect the links of a page
- Get the text of a PDF document, page by page, in the order it reads.Convert a PDF to plain text
- Work out which language a message is written in before processing it or passing it to the right person.Detect the language of a text
- Know, before calling, what a job will cost in tokens and so on the bill, to compare it with what it replaces.Estimate the cost of a model call
- Find the sums quoted in a message or a document, with their currency, however they are written.Extract amounts and currencies from a text
- Find the dates mentioned in a message or a document, whatever their format.Extract dates from free text
- Pull the number, the date and the amount due out of an invoice, whatever the supplier’s layout.Extract the fields of an invoice
- Surface the terms that characterise a document, so it can be indexed, labelled or previewed.Extract the key terms of a document
- Get an article’s text without the menu, the cookie banner, the related items and the footer.Extract the main content from a web page
- Get the rows and columns of a table printed inside a PDF — an invoice, a statement, a delivery note.Extract the tables from a PDF
- Find the records that stand for the same person or company in a file entered by several people.Find duplicate records in a file
- List the proper names mentioned in a text and tell what kind of entity each one is.Find the people and organisations named in an article
- Know how much I will sell next week, from my own sales history.Forecast next week’s sales
- Get the name, reference, brand, price and availability of an item from the page that shows it.Get the product data out of a shop page
- Separate the message that just arrived from the history copied under it.Isolate the newest reply in an email thread
- Decide, across a pile of PDFs, which pages to send to character recognition and which ones already read.Know whether a PDF needs OCR
- Prevent a phone number, an email address or an IBAN from appearing in a sent message. Not a name and not a postal address: those are personal data too, and no level on this page touches them.Mask contact details in a chat message
- Recognise that the same company appears under several spellings across two files of different origins.Match misspelled company names
- Keep an abusive comment from being published, or put it in front of a moderator before it is.Moderate user comments
- Split an address typed as a single line into number, street, complement, postcode and town.Parse an address into fields
- Show a stand-in image wherever the real one is still missing, in a mock-up or an unfinished catalogue.Produce placeholder images
- Fill a database or a mock-up with believable data, without ever pouring real customer records into it.Produce test data sets
- Order the products on a catalogue page so the most relevant ones come first.Rank products by relevance
- Get the date, the author and whatever else a document carries, before publishing it or filing it away.Read a file’s metadata
- Get the text out of a document that arrived as a PDF or an image, to index it, file it or read it back.Read the text of a scanned page
- Straighten out values from an import where “CrÃ©dit Agricole” has replaced “Crédit Agricole”.Repair text with broken encoding
- Send each incoming support ticket to the team that can answer it.Route a support ticket to the right team
- Find the right page in an internal corpus — a handbook, a knowledge base, documentation — from what a reader types into a search box.Search in your own documents
- Offer three or four other articles from the site, under the one being read, that cover the same subject.Show similar articles
- Pull the address, the date, the status code and the rest out of a log line, to count them or store them.Split log lines into fields
- Find the parcel reference in a shipping confirmation and tell which carrier it belongs to.Spot a tracking number in an email
- Get paged when a production metric leaves its usual behaviour, and not the rest of the time.Spot an anomaly in a metric
- Keep automated and promotional submissions out of a public form without blocking genuine enquiries.Spot spam in a contact form
- Add up, count or average a column of a file, without one unreadable row quietly skewing the total.Sum or average a column of data
- Get, in a few sentences, what a document too long to read says.Summarise a long document
- Attach to every article the topics it covers, taken from the newsroom’s own list.Tag articles by topic
- Make an application exist in another language, keeping the placeholders and the room available on screen.Translate interface strings
- Load a CSV file received from outside, without knowing its encoding, its delimiter or the type of its columns.Turn a messy CSV into usable data
- Get the rows and the columns of a table published in a web page, merged cells included.Turn an HTML table into data
- Check that the data received from a form is complete and coherent before storing it.Validate a form server-side
- Tell whether a number someone typed exists in a country’s numbering plan, and what kind of line it is.Validate a phone number and its country code
- Check that an incoming request matches what the API’s OpenAPI document declares for that endpoint.Validate an API payload against its contract
- Check that an address typed into a form has the shape of an address, before storing it.Validate an email address without sending a message
- Check that a JSON file received from a partner follows the agreed contract, and say exactly what is wrong with it.Validate an imported file against a schema
- Write the sales text for every item in a catalogue, starting from its characteristics.Write product descriptions
- None of those
Can the data you handle be sent to an outside provider?
Answer no if it is customer data, health data, employee data, or anything a contract forbids you to pass on.
- Yes
- No
- I don’t know
Your constraints
Do you need the same input to give exactly the same result, every time?
Yes, if a test has to pass, if someone can dispute the decision, or if two people must get the same answer.
- Yes
- No
- I don’t know
Do you need to be able to check, case by case, that the result is right?
No, if an error rate measured on a sample is enough for you.
- Yes
- No
- I don’t know
What you have
Do you already have examples of the task done by hand, with the right answer next to them?
This is what separates a rule from a trained model: with no labelled examples, there is nothing to learn from.
- Hundreds, or more
- A few dozen
- None
- I don’t know
Can someone install and run software on your own servers?
A model on your own machines costs no subscription, but it needs a machine and someone to keep it.
- Yes
- No
- I don’t know

This questionnaire needs JavaScript to chain the questions and work out the answer. Without it, start from the ten families: they sort the catalogue by what you want to do. Browse by family

Your answer

Not finding it?

The catalogue is sorted into ten families, one per verb. If you do not know what to type, start from what you want to do.

Browse by family Catalogue

Catalogue

Catalogue
Families
Latest entries
Where to start

About

About
How it works
Contribute
Contact

Legal notice

Legal notice
Privacy
Credits

Weight of this page : 79,1 Ko

No cookies, no trackers. Anonymous, cookieless analytics.

Content under CC BY 4.0 · Example code under MIT