ModelMind AI Chat Package

ModelMind started from a practical product need: an application should be able to ask questions about its own safe data without every project rebuilding the same AI chat modal, controller, prompt, database tables, and security filters. I moved that idea into a package so installation becomes a Laravel workflow instead of a copy-paste feature.

The package is published as mbs047/model-mind, uses Laravel package auto-discovery, and gives developers a first path that feels familiar: composer require, php artisan model-mind:install, php artisan migrate, then add @modelMind or the separate @modelMindStyles, @modelMindModal, and @modelMindScripts directives to a Blade layout.

I wanted the assistant to be advanced without forcing developers to touch package internals. Almost every important behavior is configurable from config/model-mind.php or environment variables: the assistant name, MBS brand mark, subtitle, launcher label, placeholder, initial message, fallback answer, tone, language behavior, OpenAI model, organization, timeout, route prefix, table prefix, session duration, retrieval limits, UI position, default questions, and public asset mode.

That matters because reusable packages live inside other teams' applications. A store may want the widget at bottom-right, an admin panel may need center-right, and a dense dashboard may want a smaller width or a custom placeholder. ModelMind treats those as first-class settings rather than hardcoded design decisions.

The assistant does not automatically read the entire application. Developers must enable each model in the models config array. For each model they can define a label, description, columns, include and exclude lists, relations, row limits, sorting, and route actions.

When columns is set to auto, ModelMind inspects the model but still filters risky fields. The security layer blocks common sensitive names such as password, remember_token, api_token, secret, recovery codes, private keys, OTP fields, card data, bank fields, salary fields, encrypted casts, hashed casts, and any hidden values the model already protects. The result is a controlled assistant context, not a raw database dump.

Config is useful, but larger Laravel apps often need behavior to live beside the model. ModelMind includes the HasModelMindContext trait so an Eloquent model can define methods such as modelMindLabel(), modelMindDescription(), modelMindContextColumns(), modelMindHiddenColumns(), modelMindContextRelations(), modelMindRouteActions(), and modelMindContextQuery().

That makes the package scalable. A product model can expose public catalog fields and storefront routes. An order model can scope the query to the current customer or hide operational notes. A content model can expose only published records. The package gives the application a central config, but it also lets each model own the details when that is the cleaner boundary.

The default modal is a polished Blade and Tailwind-friendly widget with light, dark, and auto theme support. It has a launcher, header, configured brand mark, initial assistant message, default question buttons, message history, input composer, send button, helpful and not helpful feedback states, route action buttons, and responsive positioning.

The professional part is that the default UI is not a trap. Developers can publish the Blade views, publish public CSS and JavaScript assets, or point views.modal, views.styles, and views.scripts to a completely new design. That means a team can start with @modelMind today and later ship a branded modal without editing vendor files.

A useful assistant should not only answer; it should help the user continue. ModelMind supports safe named-route actions. The developer configures approved route names and parameter mappings, for example a product action that maps products.show with ['product' => 'id'].

The package then creates buttons only from those approved definitions. It can use a static label, a label_column, or a label_template such as View {name}. This keeps links professional and safe: the AI cannot invent arbitrary URLs, but it can offer useful actions when the answer mentions configured records.

ModelMind is designed for visitors who may not ask in the same language used by the database. The assistant has language instructions that answer in the visitor's latest language, while the retrieval and route-action layers still work against the configured application data.

This became important in the Market Lane Store demo. A visitor can ask in Arabic about products whose names and descriptions are stored in English, and ModelMind can still recover the configured product route actions when the answer mentions the relevant records. The package supports multilingual conversation without requiring every database column to be translated first.

The package stores sessions and messages so the assistant does not forget the conversation after every request. It also supports browser history restoration, a recent-message prompt window, a browser-message display limit, message character limits, and MODEL_MIND_SESSION_LIFETIME_MINUTES so stale conversations can be reset after a chosen period.

For learning, ModelMind can remember assistant answers, liked answers, and manually fed texts. The learning system has minimum character limits, context limits, source metadata, and blocked patterns for secrets such as API keys, passwords, tokens, and private keys. That gives the assistant continuity without turning every chat message into uncontrolled permanent knowledge.

AI features can feel slow when every request rebuilds the same context and sends too much text to the provider. ModelMind keeps the default provider settings lean, uses a small default model option, sets connect and request timeouts, limits output tokens, caches application context, limits model rows, caps context characters, and adds question-aware retrieval so relevant records can be fetched without flooding the prompt with every configured row.

The package also includes model-mind:inspect-context and model-mind:clear-context. Those commands help developers see exactly what the assistant can read and clear cached context after changing configuration, routes, or data.

ModelMind stores its own sessions, messages, feedback, and learned knowledge. I consolidated the package tables into a single migration for fresh installs and added a configurable table prefix through MODEL_MIND_TABLE_PREFIX.

That detail matters in professional packages. Some applications want all package tables grouped with a prefix. Others need to avoid naming conflicts with existing tables. By making the prefix configurable, ModelMind can fit into more database policies without requiring custom migration edits.

The README is intentionally a starting guide: install, publish assets, configure OpenAI, render the Blade widget, enable a model, inspect context, and ask questions. The deeper documentation is split by feature so developers can jump directly to what they need.

The docs include guides for installation, Blade rendering, themes, public assets, model configuration, custom modal design, multilingual answers, retrieval, route actions, default questions, sessions, security, learning memory, feedback, table prefixes, provider settings, performance, testing, and complete simple and advanced examples. This makes the package feel closer to established Laravel ecosystem packages instead of a one-off repository.

I did not stop at source code. The repository includes a changelog, MIT license, contribution guide, security policy, issue templates, pull request template, funding metadata for sponsorship, CI workflow, Dependabot configuration, and PHPUnit tests through Orchestra Testbench.

That is part of the quality bar I wanted for ModelMind. A reusable package needs installation confidence, a clear contribution path, visible releases, a safe place to report vulnerabilities, and tests that prove the package can boot, publish, render, configure models, return route actions, restore sessions, learn from feedback, and handle multilingual answers.

The demo is installed in Market Lane Store, a sample storefront at https://mbs-sample.on-forge.com/ that showcases MBS packages. It gives ModelMind a real product catalog context, storefront routes, product detail buttons, and a UI surface where the modal can be tested like a visitor would use it.

That demo helped shape several improvements: configurable widget positions, dynamic route labels, multilingual route recovery, faster retrieval, default questions, session expiry, and public asset publishing. Instead of documenting imaginary use cases, the package was refined against a real storefront scenario.

ModelMind shows how I think about professional Laravel packages. I care about developer ergonomics, but not at the expense of security. I care about a polished default UI, but not at the expense of customization. I care about AI usefulness, but not at the expense of letting the model invent data, expose secrets, or generate unsafe links.

The result is a package that feels small to install and large in capability: one directive can render the assistant, but the system behind it includes safe Eloquent context, model traits, route actions, learning memory, feedback, multilingual behavior, configurable sessions, public assets, custom designs, tests, release notes, and a public demo.