How do you know your LLM calls actually work? What's the latency? The cost? Are the answers any good?
You need observability. Langfuse gives you that. laravel-langfuse makes it work with Laravel. And Prism is a powerful Laravel package for working with LLMs.
This repo has 9 artisan commands that show how these pieces fit together. From zero-config auto-tracing to multi-agent pipelines with scoring. Each example builds on the previous one.
- PHP 8.3+
- Docker and Docker Compose
- An Anthropic API key
git clone <your-repo-url>
cd prism
# Start Langfuse locally (takes 2-3 minutes on first run)
docker compose up -d
# Install PHP dependencies
composer install
# Configure environment
cp .env.example .env
php artisan key:generate
php artisan migrate
# Add your Anthropic key to .env
# ANTHROPIC_API_KEY=sk-ant-...The .env.example comes pre-configured with Langfuse keys that match the docker-compose.yml auto-provisioned project. No manual Langfuse setup needed.
Langfuse UI: http://localhost:3000 (login: [email protected] / password)
To stop Langfuse: docker compose down
To reset all data: docker compose down -v
| # | Command | What it demonstrates |
|---|---|---|
| 1 | php artisan example:basic-agent |
Auto-tracing with zero Langfuse code |
| 2 | php artisan example:agent-with-tools |
Tool calls appear as Langfuse spans |
| 3 | php artisan example:structured-output |
Structured JSON output in generations |
| 4 | php artisan example:streaming |
Streaming with auto-tracing |
| 5 | php artisan example:prompt-management |
Langfuse prompt management linked to generations |
| 6 | php artisan example:scoring |
Quality scores attached to traces |
| 7 | php artisan example:rag-pipeline |
Nested trace hierarchy for a RAG pipeline |
| 8 | php artisan example:multi-agent |
Multiple agents sharing a single trace |
| 9 | php artisan example:conversation |
Multi-turn conversation with session grouping |
The simplest integration. Set LANGFUSE_PRISM_ENABLED=true and every Prism call is automatically traced. No Langfuse code in your application.
In Langfuse: Auto-created trace prism-Summarizer with one generation showing model, input, output, and token usage.
Same auto-tracing, but now the agent has tools. Each tool invocation creates a span inside the trace.
In Langfuse: Trace with generation + tool-search_web and tool-calculator spans.
Agent returns structured JSON via Prism's structured output API. The structured response appears in the generation output.
In Langfuse: Generation output shows the structured JSON with sentiment, confidence, key phrases.
Streaming works identically with auto-tracing. The complete accumulated text is captured after the stream finishes.
In Langfuse: Complete generation with full text.
Create and fetch prompts from Langfuse's prompt management. Compile them with variables and link them to generations.
In Langfuse: Prompt in prompt management + trace with linked generation.
Attach numeric and categorical quality scores to traces. Useful for evaluation dashboards.
In Langfuse: Trace with generation + 3 scores (relevance, conciseness, quality).
A full Retrieval-Augmented Generation pipeline with nested spans showing each step: embedding, vector search, reranking, context assembly, and answer generation.
In Langfuse: Nested trace tree with spans, generations, events, and scores.
Two agents (Researcher and Writer) sharing a single trace. setCurrentTrace() ensures both agents' auto-traced generations nest under the same parent.
In Langfuse: Single trace with research and writing spans, each containing an auto-traced generation.
Multi-turn conversation where each turn creates its own trace, all linked by a sessionId.
In Langfuse: 3 traces grouped by session in the session view.
vendor/bin/pestTests use Langfuse::fake() to verify behavior without real API calls.
- echolabsdev/prism - A powerful Laravel package for building LLM-powered applications
- axyr/laravel-langfuse - Langfuse PHP SDK for Laravel with auto-instrumentation for Prism
Built by Martijn van Nieuwenhoven - Laravel developer specializing in AI integrations and observability tooling.
