Guide d'intégration LLM et agents IA
Cette page fournit des instructions contextuelles spécialement conçues pour être lues par des modèles de langage de grande taille (LLM) tels que GPT-4, Claude ou Gemini. Vous pouvez transmettre directement cette page à votre assistant IA pour l'aider à intégrer rapidement l'API Quote3D dans votre projet.
Invite Système
Utilisez le texte ci-dessous pour expliquer l'architecture de Quote3D à votre assistant IA ou à votre outil de génération de code (Copilot) :
You are integrating Quote3D: a REST API for instant 3D-printing quotes, with async quote jobs, webhooks, and an embeddable storefront widget. Follow this contract exactly. Where it contradicts your prior assumptions about quoting APIs, this contract wins.
== 0. TRANSPORT AND ENVELOPE ==
Base URL: https://api.quote3d.com — all paths are versioned under /v2.
Every JSON response uses one envelope: { "success": true, "data": { ... } } on success, and { "success": false, "error": "message" } on failure. File downloads and CSV exports return the raw body instead.
Branch on the HTTP status code, never on the error text — messages get reworded.
== 1. AUTHENTICATION ==
Send the token as 'Authorization: Bearer TOKEN' or 'X-API-Token: TOKEN'. Pick one style and use it consistently.
Tokens are opaque credentials — do not parse them, and do not depend on their internal structure.
Exactly one endpoint needs no authentication: the public upload route in section 2.
Token scopes: a token is either full-access or widget-scoped. A widget-scoped token only reaches what the embedded widget needs; everywhere else it is rejected with 403. Use a full-access server token for webhooks, analytics, usage and quota.
A token may also be restricted to an IP allowlist and may carry an expiry. Both failures surface as 403 and 401 respectively, not as a network error.
== 2. UPLOADING A MODEL ==
Accepted formats: STL, 3MF, OBJ. Uploads above the platform limit (50 MB by default) are rejected.
There are two upload paths. Choose deliberately.
(a) Server-side: POST /v2/file, multipart/form-data, field name 'file', authenticated. Use when the file already sits on your backend.
(b) Browser-direct, two steps: GET /v2/file/upload-id returns data.upload_id, valid for one hour. Then POST /v2/file/public/{upload_id} with multipart field 'file' and NO Authorization header. Prefer this when the shopper's browser holds the file — it keeps large uploads off your server and keeps your token out of client code.
Both paths return a file_id. Keep it; every later call is keyed on it.
File management: GET /v2/file/{file_id} downloads, DELETE /v2/file/{file_id} removes, GET /v2/user/uploads lists (newest first).
Responses carry file_path, which is the authenticated download path (/v2/file/{file_id}) — not a location on disk. Uploads are never reachable as static files, so there is no URL to link to directly.
== 3. OPTIONAL PRE-CHECK ==
POST /v2/printability/{file_id} answers synchronously with dimensions, volume, surface_area and geometric integrity (is_valid, open_edges, non_manifold_edges). Use it to reject unprintable models before spending a quote.
It needs build-volume dimensions: it reads them from the selected printer profile, or you pass x, y and z in the body. If neither is available it returns 400. The body also accepts technology.
== 4. QUOTING IS ASYNCHRONOUS — THIS IS THE PART MOST INTEGRATIONS GET WRONG ==
POST /v2/file/quote/{file_id} does NOT return a price. It answers 202 Accepted with { jobId, status: 'queued', statusUrl, estimatedTime, createdAt }.
POST /v2/file/quote/{file_id}/async is the same endpoint under a second name. Do not build different logic for the two.
Request body, all optional: technology ('FDM' | 'SLA' | 'SLS'; 'RESIN' is an alias for 'SLA'), printer_id, quantity, and the three config objects printer_config, material_config and quote_config.
Anything you omit is resolved for you: request value first, then the user's dashboard profile, then the global profile. So the minimal useful body is often {} or { "quantity": 2 }. Do not invent required fields, and do not send a wall of parameters an integrator would rather configure once in their dashboard.
Then poll GET /v2/jobs/{job_id} and read data.status. Statuses are LOWERCASE: 'queued', 'processing', 'completed', 'failed', 'cancelled'. Terminal statuses are 'completed', 'failed' and 'cancelled'.
A failed calculation still answers HTTP 200, with status 'failed' and data.error = { message, code }. Never treat 200 as success — inspect data.status. This is the single most common bug in generated Quote3D clients.
While processing you may also read data.progress and data.estimatedTimeRemaining. On completion the quote payload is at data.result.
Poll with a bounded loop: a fixed 2-3 second interval or exponential backoff, plus a hard attempt cap and a timeout path. Never poll without a ceiling.
== 5. READING STORED QUOTES ==
GET /v2/quotes lists them, GET /v2/quotes/{quote_id} returns one, DELETE /v2/quotes/{quote_id} removes one.
Pagination: limit and offset query parameters, default 50, maximum 100. The response carries a pagination object with total, limit, offset, has_more, page and total_pages — use has_more rather than computing the end yourself.
Sorting is one repeatable parameter in field:direction form, e.g. ?sort=created_at:desc.
Caveat: the detailed 'result' block on a stored quote is a stored payload, and older or partial records fall back to a smaller shape holding only pricing.total, pricing.currency, timeEstimation and modelInfo. Read defensively with optional chaining instead of assuming the rich shape.
== 6. WEBHOOKS — THE PRODUCTION PATTERN ==
Prefer webhooks over polling for anything long-lived.
Manage them with POST /v2/webhooks (body: { url, events }), GET /v2/webhooks, GET/PUT/DELETE /v2/webhooks/{webhook_id}, and POST /v2/webhooks/{webhook_id}/deliveries/{delivery_id}/resend to redeliver.
The signing secret is returned ONLY in the create response. Store it immediately; it cannot be read back.
Event catalogue, exhaustively: 'quote.completed', 'quote.failed', 'file.uploaded', 'file.deleted', 'job.status_changed', 'widget.added_to_cart'.
Verification: compute HMAC-SHA256 over the RAW request body with the secret and compare against the 'X-Webhook-Signature' header, whose value is the literal prefix 'sha256=' followed by the hex digest. Use a timing-safe comparison. The event name also arrives in 'X-Webhook-Event'.
Read the raw body before any JSON body parser touches it, or the signature will never match.
Deliveries retry, so handlers must be idempotent — key on the event id or the quote id, and make repeat delivery a no-op.
== 7. THE EMBEDDABLE WIDGET ==
Load /js/quote3d-embed.js, then Quote3D.init('#root', { token, theme, color, locale, redirectUrl, quoteId, onResult, onAddToCart }). There is also a baseUrl option, which defaults to the origin serving the script.
The SDK builds an iframe URL and maps redirectUrl to the query parameter 'redirect_url'.
Give the widget a widget-scoped token, never a full-access one — it is visible in client code.
The iframe posts three message types to the host: 'QUOTE3D_RESULT', 'QUOTE3D_ADD_TO_CART' and 'QUOTE3D_RESIZE'. Handle resize by setting the iframe height; ignoring it leaves the widget clipped.
Payload contract: quoteId, price, unitPrice, currency, material, color, technology, quantity, fileName, weight, filamentWeight, estimatedTime, dimensions, and print settings such as layerHeight, infill, infillPattern, walls, postProcessing, hollowing.
The total is 'price'. There is NO 'totalPrice' field — reading it yields undefined and silently breaks carts.
The add-to-cart payload adds thumbnail, thumbnailUrl, thumbnailBase64 and addedAt. Keep handling both weight and filamentWeight for backward compatibility.
thumbnailUrl is an absolute, signed URL that renders from any origin — store it as given and never rewrite or re-host it, or the image stops resolving.
Do not call the widget's own internal routes from your code. To react to a shopper adding a configured part to the cart, subscribe to the 'widget.added_to_cart' webhook.
== 8. ACCOUNT, LIMITS AND REPORTING ==
GET /v2/user — account, plan and monthly allowances: quotes_used, quotes_limit, storage_used, storage_limit, files, days_till_reset, reset_date.
GET /v2/quota — rate-limit allowances (global plus a per-endpoint breakdown with remaining and reset times) and request counts for today, this month, this year and all time.
GET /v2/usage — usage analytics including per-endpoint and per-material statistics.
GET /v2/analytics/quotes (period=day|week|month|year|custom, with date_from and date_to when custom), /v2/analytics/popular (limit, default 10), /v2/analytics/cost-trends (group_by=day|week|month), /v2/analytics/export (format=json|csv).
== 9. ERRORS AND RESILIENCE ==
400 — invalid body or query, unsupported format, or a model that does not fit the selected printer.
401 — token missing, malformed, expired or revoked.
403 — valid token, not permitted: scope restriction or IP allowlist.
404 — the file, quote, job or webhook does not exist or belongs to another account.
429 — rate limited. Read the 'Retry-After' header and back off exponentially.
500 — retry with backoff; if it persists, surface the timestamp to the user.
Rate-limit state also arrives on successful responses via 'X-RateLimit-Limit', 'X-RateLimit-Remaining', 'X-RateLimit-Reset' and 'X-RateLimit-Window'. Limits are configurable per deployment, so read them from these headers or from GET /v2/quota — never hardcode a number.
== 10. DO NOT DO THESE ==
- Do not expect a price from the quote endpoint, and do not add a 'wait' that assumes the job finished.
- Do not compare job status against uppercase strings.
- Do not treat HTTP 200 on the jobs endpoint as success.
- Do not read 'totalPrice' from a widget payload.
- Do not put a full-access token in browser code, an iframe URL, a query string or a redirect payload.
- Do not JSON-parse the webhook body before computing its signature.
- Do not hardcode rate limits, quota numbers or the file-size cap.
- Do not invent endpoints, fields or query parameters. If this contract does not name it, ask instead of guessing.
== WHAT TO PRODUCE ==
- Complete upload -> optional printability check -> async quote -> bounded polling -> result flows.
- Widget embed code with resize handling, redirect handling and postMessage listeners.
- Add-to-cart bridges that preserve the payload fields above and read 'price'.
- Webhook handlers with raw-body signature verification, timing-safe comparison and idempotent processing.
- Error handling that distinguishes 400/401/403/404/429/500 and honours Retry-After.Contexte LLM pour l'intégration de widgets
Vous pouvez demander à votre assistant LLM d'effectuer les tâches liées aux widgets suivantes :
- Configuration du SDK JS : Générez une configuration Quote3D.init qui utilise un token, une locale, un thème, une couleur, un quoteId optionnel et un redirectUrl.
- Écoute d'événements : Écrivez des écouteurs de page hôte pour les événements QUOTE3D_RESULT, QUOTE3D_ADD_TO_CART et QUOTE3D_RESIZE.
- Mappage de la charge utile : conservez les champs renvoyés tels que quoteId, price, unitPrice, currency, technology, quantity, thumbnailUrl, fileName, weight et filamentWeight. Notez que le total est price, et non totalPrice.
Logique de traitement des Webhooks
Fournissez ces détails pour que votre bot écrive le code de sécurité et de gestion des événements des webhooks :
- Vérification de signature : écrivez des fonctions qui vérifient l'en-tête X-Webhook-Signature via HMAC-SHA256 sur le corps brut de la requête. La valeur de l'en-tête est le préfixe littéral sha256= suivi du condensé hexadécimal : comparez-la avec une comparaison à temps constant plutôt qu'avec une simple égalité.
- Gestion des événements : traitez l'ensemble du catalogue d'événements - quote.completed, quote.failed, file.uploaded, file.deleted, job.status_changed et widget.added_to_cart - avec des gestionnaires idempotents.
Meilleures pratiques pour les agents
Lors de la rédaction du code d'intégration pour Quote3D, gérez toujours les éléments suivants :
- Implémentez un sondage asynchrone pour la génération de devis. Ne vous attendez pas à des retours de prix immédiats lors du POST initial.
- Utilisez un retour exponentiel ou une boucle de délai fixe de 2 à 3 secondes lors du sondage
/v2/jobs/{job_id}. - Conservez les deux
weight&filamentWeightdans les métadonnées du panier ou de la commande, pour rester compatible avec les charges utiles et plugins actuels du widget. - Privilégiez les webhooks pour les intégrations en production et rendez les gestionnaires idempotents car des tentatives peuvent se produire.
Résumé des points de terminaison critiques
Un index de référence rapide des points de terminaison les plus utilisés pour une implémentation de base :
Flux d'intégration de bout en bout
L'ordre attendu par l'API et ce que chaque étape transmet à la suivante. Demandez à votre assistant d'implémenter ces six étapes plutôt que de décrire un point de terminaison isolé — la plupart des bogues d'intégration viennent d'une étape sautée ou d'un enchaînement mal supposé.
- Créez un jeton dans le tableau de bord et choisissez sa portée : accès complet pour votre serveur, portée widget pour tout ce qui est intégré à une boutique.
- Transmettez le modèle à Quote3D et conservez le file_id renvoyé. Envoyez-le depuis votre backend avec POST /v2/file, ou laissez le navigateur l'envoyer directement via GET /v2/file/upload-id puis POST /v2/file/public/{upload_id} — la seconde voie évite les gros fichiers sur votre serveur et votre jeton dans le code client.
- Vérifiez éventuellement le modèle en amont avec POST /v2/printability/{file_id} afin de rejeter une géométrie non imprimable avant qu'elle ne consomme un devis.
- Lancez le devis avec POST /v2/file/quote/{file_id}. Il répond 202 avec un jobId, pas un prix. N'envoyez que les paramètres qui varient d'un devis à l'autre ; le reste provient des profils du tableau de bord.
- Attendez le résultat. En production, abonnez-vous aux webhooks quote.completed et quote.failed. Pour les scripts et prototypes, interrogez GET /v2/jobs/{job_id} dans une boucle bornée jusqu'à un statut completed, failed ou cancelled.
- Lisez le résultat depuis la charge utile du travail, ou plus tard via GET /v2/quotes et GET /v2/quotes/{quote_id}. Gérez les 429 en respectant partout l'en-tête Retry-After.
Les trois erreurs à l'origine de la plupart des intégrations défaillantes : attendre un prix du point de terminaison de devis, comparer le statut du travail à des chaînes en majuscules, et considérer un HTTP 200 sur le point de terminaison des travaux comme un succès alors que le travail a échoué. Le prompt système ci-dessus énonce ces trois points explicitement pour que votre assistant ne les reproduise pas.