SourceOpen — LLM chatbot and QR attendance prototypes in Laravel

Two questions were worth answering with working code rather than reading: how a conventional rule-based chatbot framework can hand off to an LLM for anything it does not recognise, while keeping the model on-topic for a specific organisation; and how a QR-code check-in flow can be built quickly for events and classes without a mobile app.

Personal prototype repository · PHP / Laravel

Approach

• Use BotMan's web driver for deterministic intents (greetings, capturing a name) and register a fallback that calls a hosted open model with an organisation-specific system prompt.
• Keep short-term conversation memory per sender in the Laravel cache through a BotMan middleware, capped in size and expiring after a week.
• For attendance, generate an SVG QR code per person and event, and expose a small API endpoint that a scanner posts to; the endpoint forwards the check-in or check-out to an existing backend service.

A single Laravel application with two independent features. The chatbot route builds a BotMan instance per request, matches known intents, and otherwise sends the message and system prompt to the model provider. The attendance flow renders QR codes server-side and proxies scan events to an external API.

• 01BotMan 2.8 with the web driver; intents registered with `hears()` and a `fallback()` that posts the user's message to the Together AI chat-completions API using `meta-llama/Llama-3.3-70B-Instruct-Turbo`.
• 02A `MessageHistoryMiddleware` implementing BotMan's `Received` interface stores timestamped messages per sender in the Laravel cache, truncated to 2,000 characters.
• 03Separate bot endpoints with different system prompts, showing how the same pipeline is reused for different organisations.
• 04QR codes generated with simple-qrcode as SVG, encoding a JSON payload of user and event identifiers; a Sanctum-ready API route receives scans and forwards them with cURL to the existing backend.
• 05An `attendances` table with a unique constraint on student, event and date to prevent duplicate records.

The LLM is used as a fallback behind deterministic intents — the pattern that keeps a chatbot predictable for known questions while still handling open-ended ones. The system prompt constrains the model to the organisation's context and instructs it to decline unrelated topics.

• Laravel 11
• PHP 8.2
• BotMan
• Together AI (Llama 3.3 70B)
• Laravel Sanctum
• simple-qrcode
• MySQL

Outcome to be added. Results are only published here when they can be stated factually — no estimated metrics.

• Rule-based first, model second: deterministic handling for the common cases keeps behaviour predictable and costs down.
• Conversation memory needs a size cap and an expiry from day one.
• Prototype code that disables TLS verification or truncates tables via a GET route is fine for a lab — and exactly what must be removed before anything goes to production.