Search
Enter at least two characters without personal information.

    LibraryArticles & FAQ Your projectRequest a quote
    LanguageEN
    Web resources

    Artificial intelligence

    AI resources mix model documentation, services and communities. For professional use, separate demonstrated capability from data handling, output verification and operating cost.

    Prepare a small set of real tasks with expected answers, allowed sources and failure cases. Compare outputs, observed cost and data handling before integrating a model.

    AI-generated illustration

    Editorial method

    A directory made to be useful, not merely long.

    Selection prioritises recognition within the field, professional use and reference status. Highly specialised documentation entries are consolidated under their main resource.

    01

    Official foundations

    Standards bodies, official documentation and public authorities are used as primary reference points.

    02

    Professional tools

    Established services are grouped by actual use: building, hosting, measuring, securing, publishing or promoting a website.

    03

    Independent reading

    Inclusion is editorial and unpaid. A listing does not imply a partnership or guarantee.

    Before choosing

    Compare on three practical questions

    Artificial intelligence
    CriterionQuestion to resolve
    Submitted dataWhich data leaves the project, for what purpose and for how long?
    EvaluationWhich real examples will expose errors, omissions and invented results?
    ReviewWho reviews the response and how can work continue if the service fails?

    Keep a short decision record: requirement, source, observed result and unresolved point. Test the same task with each candidate instead of comparing marketing lists of features.

    Search and filter

    Official starting points for this topic

    Compare their role and documentation before selecting a tool. Inclusion is not a ranking.

    13 resources displayed

    The date applies to the linked document, not to a review of every service feature.

    Artificial intelligence

    Mistral AI

    French company developing generative AI models and services.

    Mistral Docs · Source consulted

    Use it for : Compare deployment choices with data and maintenance requirements.

    Documented characteristics
    Endpoint
    POST /v1/chat/completions
    Format
    application/json
    Authentication
    Bearer

    Verification path

    1. Define source and fields
    2. Separate format from accuracy
    3. Validate every value to publish

    Learning example: adapt files, identifiers and conditions to a test copy before interpreting the result.

    Practical example
    Fictional example: turn an event note into a structured record. Define expected fields and keep null for a date absent from the source instead of inventing one. Use a demonstration document without private data.
    Useful check
    The documentation distinguishes JSON output and schema-constrained output. A valid structure checks form, not the truth of names, dates or quotations. Compare every value with the source passage and handle missing information explicitly.
    Interpretation and next step
    Record model, parameters and document version during the trial. A plausible or reproducible response is not evidence. Have the content owner validate publication elements, particularly after model or instruction changes.

    Official documentation ·

    Official documentation · Mistral · Structured outputs

    FranceInstitution / standard
    mistral.aiVisit official website
    Artificial intelligence

    Hugging Face

    Community platform for AI models, datasets and demos.

    Hugging Face · Source consulted

    Use it for : Explore models and datasets while checking their licences and limitations.

    Documented characteristics
    API
    Hub OpenAPI
    Catalogues
    Models · Datasets · Spaces
    Documentation
    Model cards · Dataset cards
    Format
    Markdown · YAML
    Configuration
    README.md

    Learning example: adapt files, identifiers and conditions to a test copy before interpreting the result.

    Practical example
    To compare two models intended for the same task, start with their README.md model cards. These combine Markdown text with YAML metadata. Record the pipeline_tag task, described uses and stated limitations; a text-generation model does not automatically meet every requirement of an assistant.
    Useful check
    Check the license field and the licence text it references. Also identify the listed datasets and the conditions applying to the resources used. An incomplete card leaves questions open: do not turn missing information into permission or a favourable evaluation result.
    Interpretation and next step
    Keep the repository identifier and revision with the test files and settings. A main branch can change, so two tests using the same model name are not necessarily comparable. Evaluate responses on representative cases and distinguish the publisher’s results from those you actually reproduced.

    Official documentation ·

    Official documentation · Hugging Face — Model cards

    Official documentation · Hugging Face Hub — Download files at a specific revision

    InternationalCommunity / learning
    huggingface.coVisit official website
    Artificial intelligence

    Cursor

    AI-assisted code editor and programming agent.

    Cursor · Source consulted

    Use it for : Evaluate a representative task, review proposed changes and check transmitted data before authorizing an agent on a sensitive repository.

    Cursor — Cursor brand guidelines
    InternationalTool / software
    cursor.comVisit official website
    Choose a starting point

    What each category helps you decide.

    These examples illustrate uses, not rankings or endorsements. Verify each provider’s current documentation and terms before a decision.

    Artificial intelligence

    AI resources mix model documentation, services and communities. For professional use, separate demonstrated capability from data handling, output verification and operating cost.

    Questions to check
    • Test representative tasks and inspect failures.
    • Review data-handling terms before sending sensitive material.
    • Check an answer against its source and test a case with missing required information.

    A practical use : Prepare a small set of real tasks with expected answers, allowed sources and failure cases. Compare outputs, observed cost and data handling before integrating a model.

    What this cannot prove : One demo or fluent answer does not establish reliability on difficult cases. Public facts repeated by an assistant still need checking.

    Examples : OpenAI Developers · Hugging Face

    Related guide : Verify sources and evidence in web content

    Explore troubleshooting situations

    Troubleshooting methods

    An AI trial becomes informative when it preserves failures as well as good answers. Prepare tasks with verifiable results, define allowed sources and keep a human recovery path. These situations do not rank suppliers; they describe checks for your own use.

    A convincing answer cites a source that cannot be found

    What you observe
    The answer gives a credible-looking title, link or quotation, but the document cannot be found or does not contain the claimed information.
    Before drawing a conclusion
    A citation establishes neither the existence nor the content of a source. Verification must examine an accessible passage.
    Verification step
    Open the document, locate the information and record the result. Include a question without an available answer in your trial set and check that the process can acknowledge the gap before publication.

    An isolated success does not repeat

    What you observe
    The same task sometimes produces a correct result and sometimes an omission, ambiguous wording or an invented detail.
    Before drawing a conclusion
    The best observed example does not describe error frequency. A comparison must preserve conditions and all outputs from its planned trial.
    Verification step
    Apply the same criteria to several representative examples, including difficult cases. Classify errors by consequence and document required human review, without reporting a rate from a sample selected after the fact.

    A processed document contains unexpected instructions

    What you observe
    A text to summarise or a retrieved page asks to change rules, execute an action or send information elsewhere.
    Before drawing a conclusion
    Source content may be untrusted. An instruction inside it must not automatically become permission to act.
    Verification step
    Add a demonstration document containing a conflicting instruction. Check that processing stays within the authorised task and that external actions retain their controls. Use no sensitive data in this trial.
    Another case to examineThe summary dropped the condition that changes the decision
    Situation
    A summary says a feature is available, although the document restricts it to certain accounts. The sentence seems correct when its context is omitted.
    Before drawing a conclusion
    A genuine source can be cited yet summarised inaccurately. Finding the link alone does not verify the claim.
    Practical check
    Compare the answer with the whole passage, including exceptions and notes. Rewrite the conclusion with its conditions and test a question explicitly addressing the excluded case.

    Further documentation

    Keep useful data when changing tools

    Before making a resource part of website operations, plan for updates and an exit too. These checks concern your trial results without assuming features or prices of a particular plan.

    Keep the evaluation set

    Retain authorised cases, expected answers and scoring method. A model change should be comparable on the same tasks.

    Separate model and journey

    Identify output formats, limits and errors handled by the site. A replacement provider can return a different format for the same instruction.

    Check the exit

    Examine applicable terms for submitted data and deletion. Keep validated content in your editorial system rather than solely in conversation history.

    Transparency

    Editorial scope and third-party websites

    This non-exhaustive directory had its data consistency and links checked on 2 October 2026. Some websites restrict automated checks. It contains external links to independent websites. Their availability, content, prices, cookies and terms may change. Française du Numérique does not control them and receives no commission for their inclusion.

    Directory questions

    Selection, independence and updates.

    Entries describe resources from their publishers’ pages and provide a source, category, purpose and checking date. The selection covers different Web tasks rather than ranking every available provider. Open the cited source to check current scope, availability and conditions before relying on a service.

    No. Inclusion identifies a resource to examine; it does not certify every feature, price, policy or result. Compare the resource with your actual task, country, language and technical requirements. Confirm current conditions with its publisher before committing time, data or money.

    Yes. Search the displayed catalogue, combine the available filters and clear them to return to the full list. Every entered word must match the indexed text. The visible result count describes this selection, not the entire Web. Individual category pages and direct links remain available without JavaScript.

    Use authorised public examples or clearly labelled demonstration data. Preserve task difficulty while removing identities, credentials and confidential information. Check service terms before any transfer and document which sources were actually used.

    Also examine omissions, ambiguous cases, interruptions and the consequences of an incorrect result. Define who can review, correct and resume work. Automation should remain proportionate to what representative cases have actually tested.

    Compare using observable evidence

    Before choosing a resource, run a limited trial using the same data and task. Keep the outcome and its limitations rather than a ranking based on familiarity.

    Use a small set of representative questions with verifiable answers. Record model, date, instructions, errors and unanswered cases. Compare accuracy and usefulness on that set; fluent wording or an isolated result does not demonstrate general reliability.

    Use precise terms

    Two distinctions before comparing resources

    A cited source and a numerical summary each need an appropriate check.

    Documented reference points for this topic

    Documented reference points for this topic
    ReferenceValue or syntaxScope
    RFC 8259 · JSON
    application/json · UTF-8Standard JSON permits neither comments nor trailing commas; strings use double quotes.

    Dates indicate when references were consulted. These facts are neither a site audit nor an overall compliance statement.