DEV3 Promptly – AI Content & Chatbot SaaS Laravel Script with AI Writer, Document Chat & Credit Plans

The engineering behind Promptly, the AI SaaS of the DEV3 Laravel suite: Laravel 12 and Filament 4 on the shared DEV3 core and SaaS packages, a provider-agnostic DEV3 AI pipeline with an offline demo model, streaming without websockets, a credit ledger, cited retrieval and an embeddable widget.

The problem it solves

An AI SaaS has a cost problem that most SaaS products don't: every click spends real money with an AI provider, and the customer's plan has to cover it. Usage must be measured in tokens, turned into something customers understand, checked before the request runs and charged after it, without ever letting a busy workspace go below zero or a slow request charge twice. Buyers of a script also want a free choice of AI vendor, and they need to install, demo and test the product without an API key or a bill, so the AI layer had to be provider-agnostic, with an offline model good enough to show every feature. Answers about a customer's own documents are only useful if they can be checked, so each one needed to point at the page and section it came from. A website chatbot runs on other people's sites, where it must not break their CSS, must not answer for sites it wasn't installed on, and has to hand a visitor to a person when the bot can't help. Streaming had to work on shared hosting with no websockets, Node.js or Redis. And all of it had to share the workspaces, plans, invoices and MRR of the DEV3 SaaS package already used by Slotwise, and install from a zip through a web wizard.

Architecture and approach

Promptly is a Laravel 12 application with two Filament 4 panels on one user table: /app/{workspace} for customers, with Filament tenancy so each URL belongs to one workspace, and /admin for the platform owner and the support role. The public website, the live demo box and the chatbot widget are Blade with a small vanilla-JS layer, which keeps the public page at 29.0 KB of gzipped JS and CSS.

Two shared DEV3 packages

• The core. dev3/laravel-core provides the installer, typed and encrypted settings, branding, the Blade UI kit, motion and 3D helpers, demo mode, roles and permissions, the audit log, payment gateways, backups, system health, Sanctum tokens, English and Arabic with RTL, and the AI contracts: an AiClient interface with streaming and token counting, and a CreditLedger extension point.
• The SaaS layer. dev3/laravel-saas adds single-database tenancy by team_id with a global scope, invitations, plans with limits and metered usage, trials, subscriptions, PDF invoices and MRR metrics. In Promptly a Workspace extends the package's Team with credit columns, a brand voice and default tone and language, so documents, sources, chatbots, conversations, leads and ledger rows are isolated per workspace without a line of filtering in the product code.

DEV3 AI: one pipeline, any provider

• Every request takes the same path. The writer, the editor actions, document chat, chatbot replies and the REST API all call one service, PromptlyAi: moderation, then a credit check against an estimate (the prompt's tokens plus the expected answer), then the driver, then a usage log, then the debit for the tokens actually used.
• Customers see models, not vendors. Each model row ("DEV3 AI Swift", "Balanced", "Deep Context", "Pro") maps to a driver and an exact provider model id with its own credit multiplier. A row on "Default provider" follows Settings, so switching vendor is one setting, and the UI, the widget and the website only ever say DEV3 AI.
• Drivers are thin HTTP clients. OpenAI-compatible, Anthropic and Google drivers implement the core contract with streaming, and buyers can register their own. Keys are stored encrypted in Settings, and the OpenAI-compatible base URL is a setting too, so any Chat Completions-compatible endpoint, hosted or self-hosted, works without code.
• An offline demo model that shows every feature. The default fake driver makes no network calls. A seeded content engine writes template-aware drafts (blog posts, ad sets, emails, FAQs and more) in English or Arabic, the same input always gives the same draft, a rewriter handles the editor actions, and a composer builds cited answers from the retrieved chunks. It is what the demo, the tests and every screenshot use, so no real AI key is ever needed to develop, test or seed Promptly.

Credits: an append-only ledger

• One formula. Credits are ceil(tokens ÷ tokens per credit × feature multiplier × model multiplier), with a minimum of 1. The defaults are 10 tokens per credit; writer 1.0, editor actions 0.8, document chat 1.0, chatbot replies 0.8 and indexing 0.1; and model multipliers from ×0.5 to ×2. All of them are admin settings.
• Rows, not a running total you can't explain. WorkspaceCreditLedger implements the core's ledger contract. Every grant, renewal, debit and adjustment is a row with its tokens, feature, model, member and the balance after it. A debit locks the workspace row inside a transaction and charges at most what is left, so concurrent requests never push a balance below zero.
• Periods and alerts. An hourly promptly:renew-credits job resets each workspace whose renewal date has passed to its plan allowance, so unused credits expire, and records the change as a renewal row. Low-balance emails and in-app notices fire once at 20% and once at 5%, and a new grant re-arms them.
• A hard stop with a way out. When the estimate doesn't fit, the request never reaches the provider: the writer shows an "Upgrade plan" link, the API returns 402, and a chatbot switches to a fallback that flags the conversation for a person and asks the visitor for their details. When a subscription starts or a plan changes, an event listener moves the allowance to the new plan at once: an upgrade tops the balance up to the new allowance and keeps anything above it, and a downgrade keeps min(remaining, new allowance), each as a ledger row. A super admin can also grant credits, which writes a topup row and an audit entry. Each debit is also recorded as metered usage in the SaaS package, which drives the meters on the Billing page and in the admin tenant view.

Plans, roles and private files, enforced on the server

• Plan features are checks, not labels. Lead capture, REST API access, premium DEV3 AI models and removing the "Powered by" badge are per-plan features checked in the widget endpoints, the inbox, the writer, the API and the chatbot form. A locked feature shows an "Upgrade plan" link in the app, and the API answers 403 with a plan_feature code.
• Read-only means read-only. Workspace roles map to abilities (content.write, chatbots.manage, settings.manage) enforced by model policies, page and action guards, the streaming endpoints and the API. A viewer can read everything but can't run DEV3 AI, edit or reply; owners and admins can do everything.
• Private knowledge files. Uploads go to a private knowledge disk as workspace-<id>/<random name>, with the original name kept in the database. The only way to read one is an authorised download route that checks membership of the source's workspace; other workspaces get a 404. An upgrade migration moves 1.0.0 files off the public disk, and the core's backups include the private folder.
• Two-factor sign-in from the core. Filament's multi-factor authentication is wired once in the DEV3 core: an authenticator app with 8 recovery codes, or email codes. /admin honours a "Require 2FA for staff" setting, workspaces keep it optional, a super admin can reset it (also php artisan dev3:2fa-reset), and one-click demo logins never see a challenge.

Streaming without websockets

• Server-sent events over a POST. The writer, the editor and document chat post to /app/stream/* and read a text/event-stream response with chunk, done and error events. The response turns off PHP output buffering and compression, sends X-Accel-Buffering: no and a 2 KB padding comment so proxies flush early. It works with PHP-FPM on shared hosting, with no websockets, Node.js or Redis; 1.0.1 even drops the unused Reverb package.
• The same shape for every driver. Real drivers stream provider tokens through a callback. The demo model writes its answer and replays it word by word at a short, capped pace, so the demo feels like the real thing. The neural orb listens for promptly:orb browser events and speeds up while tokens arrive.

Retrieval with citations

• Extraction. PDFs are read page by page, DOCX files keep their headings, and imported web pages lose their scripts, navigation and footers.
• Chunking that remembers where it came from. A paragraph-aware chunker with a sentence fallback cuts about 900 characters with 150 characters of overlap, keeps the nearest heading, and never lets a chunk span two pages, so a citation's page number is exact. The heading is embedded with the chunk's text.
• Embeddings with no API by default. The built-in provider is 256-dimension signed feature hashing of words and word pairs, with stop words removed and L2 normalisation: deterministic, free and good at the keyword-heavy questions people ask about price lists and policies. An OpenAI-compatible embeddings endpoint can replace it. Indexing is charged at the low indexing rate.
• Answers that show their sources. The question is embedded, the chunks of the selected sources are scored by cosine similarity, and the best four above a 0.08 floor become numbered context that DEV3 AI must cite as [n]. Afterwards Promptly keeps only the citations the answer actually used and stores each one with its source, page, section and excerpt, so the reader can check the claim.

A chatbot widget that is safe to embed

• Isolated on the host page. One script tag loads widget.js, which renders inside a Shadow DOM, so the host site's CSS can't reach it and its CSS can't leak out. It reads its colours, texts and suggested questions from the bot's config.
• Bound to its sites and its visitor. Each request's Origin (or Referer) must match the bot's allowed domains, subdomains included, or it gets a 403. Starting a chat issues a visitor token, a base64url payload of the conversation, bot, visitor and expiry signed with HMAC-SHA256 using the app key and the bot's public key, valid for 24 hours. Messages, polling, leads and ratings all need it, so there is no cookie on third-party sites and one visitor can't read another's conversation. The endpoints are rate-limited per bot and IP.
• From bot to person. When a visitor asks for a human, the bot answers that the team has been told and flags the conversation. The team's inbox refreshes on its own, an agent's reply clears the flag, and the widget picks it up on its next poll. Lead capture asks for the chosen fields after a set number of questions or at hand-over, creates the lead and notifies the owner by email and in the app.

Moderation and abuse flags

• Every prompt and widget message passes a moderation step before any credit is spent. Blocked terms are refused with a high-severity flag (the API answers 422). Prompt-injection phrases, card or ID numbers (stored masked), messages with many links and per-workspace rate spikes are flagged but allowed, so false positives don't block customers. Review happens in the admin's Abuse flags queue, where taking action suspends the workspace and stops its AI and its chatbots.

An installer and demo data made by the real services

• Installs like a script. If .env is missing the app creates it, then redirects to a 5-step /install wizard (requirements, database, demo or empty data, admin account, finish); php artisan dev3:install does the same over SSH. vendor/ and public/build ship in the zip, the queue defaults to sync, and the scheduler also drains queued indexing jobs on hosts without a worker.
• A believable demo. The seeders build "Brightpath Media", a content agency on the Pro plan with five members and a brand voice, six more named workspaces from a travel company to a low-credit analytics start-up, and 16 smaller ones on every plan. Months of writer runs, editor actions, document chats, chatbot conversations, hand-overs and leads all go through the real AI pipeline, ledger, ingestor and bot service with the demo model, so balances, usage charts and MRR reconcile. The knowledge-base PDFs are generated by code. The web installer seeds about 35% of that history to stay inside shared-hosting request limits.

The 3D and motion approach

• A neural orb in the hero. About 150 nodes on a sphere, linked to their nearest neighbours, with an inner core, a token ring and pulses that travel along the links. Typing in the live demo box sends a ripple; streaming makes the pulses race and the core brighten, then the orb settles.
• Token particles on the dashboard. The workspace's daily token usage becomes a radial chart of particle columns around a glowing core, with today's column brightest and a spiral of particles streaming inward at a speed set by recent usage.
• Cheap by default. three.js loads lazily as its own 187 KB chunk, the device pixel ratio is capped, scenes pause off-screen, reduced motion is respected, and SVG posters show without WebGL or when 3D is switched off in Settings → Branding. Reveals and counters come from the core's data-attribute motion.js with Lenis smooth scrolling, not GSAP.

Quality gates and results

Promptly 1.0.1 passed all seven release gates of the DEV3 Laravel suite in one run against the packaged build:

• Static analysis: composer validate, php -l on 297 files, Pint clean and Larastan level 6 with 0 errors, with the core and SaaS packages included.
• Tests: 132 Pest tests with 708 assertions, including required money-path suites for the credit ledger, metering and plan limits, RAG chunking and citations, and widget security, plus suites for private knowledge storage, plan features, the read-only viewer role, credit top-ups and two-factor sign-in. Every test runs on the offline demo model.
• Front-end budget: 11.9 KB of JS and 17.1 KB of CSS gzipped on the public page (29.0 KB against a 180 KB budget), with the 187 KB three.js chunk loaded only when a scene needs it.
• Databases: the demo and empty seeders run on both MySQL 8 and SQLite. The demo seed writes 498 generations, 987 ledger rows, 935 AI usage records, 112 documents, 128 document-chat messages, 276 chatbot conversations with 1,116 messages, and 87 subscription payments.
• Browser checks: 117 pages in headless Chrome with 0 console errors, plus 4 happy paths: sign-up to a streamed writer result saved into the editor, a document-chat answer with citations, a widget chat that becomes a lead and gets a reply from the inbox, and a low-credit workspace reaching the upgrade prompt.
• Installer: the release zip unzipped into a clean folder, installed through the /install wizard on MySQL with demo data, then signed in as the admin.
• Secret scan: 19,120 files in the 25 MB release zip, clean.

The 118 screenshots and the two video tours were captured from the seeded demo across both panels, the public site and the widget, in light and dark mode, Arabic, tablet and phone. They follow the main flows step by step: a template from brief to streamed draft to document, a question answered with citations, a chatbot from training to embed code to a captured lead, and a workspace running low on credits and upgrading. Every answer in them comes from the offline demo model.

Version 1.0.1 closed the gaps we had listed for 1.0.0: two-factor sign-in, private knowledge uploads, server-side plan features, a read-only viewer role and super-admin credit top-ups.

The remaining known gaps are listed openly:

• The OpenAI-compatible, Anthropic and Google drivers and the Stripe, PayPal and Razorpay gateways have not been run against live accounts.
• The demo model writes English and Arabic only, and the web installer seeds a lighter demo than the command line.
• The Billing page's usage meter counts calendar months while the ledger follows each workspace's renewal date.

Tech stack

Inside the build

Home page with the 3D neural orb — A glowing network of synapses reacts to the cursor in a GPU-rendered hero, with a clear sign-up call to action.

Live “try DEV3 AI” box — Visitors type a request and watch the answer stream in word by word while the neural orb pulses with every token.