Upgrade Guide

Updating Dependencies

You should update the following dependencies in your application's composer.json file:

• inspector-apm/neuron-ai to ^2.0

High Impact Changes

Workflow

The Workflow component was completely re-architected. The Edge class was completely removed and now the orchestration system is event-driven relying only on node definition.

It also supports real-time streaming.

We strongly recommend to use the V2 documentation to understand the new system and move your previously created Workflow to the new architecture.

Workflow Documentation

Low Impact

RAG Retrieval component

In the previous version the RAG component could interact with a vector store and an embeddings provider, but there was no way to customize this behavior. Recently many different retrieval techniques emerged trying to increase the accuracy of a RAG system.

Neuron RAG now has a separate retrieval component that allwos you to implement different strategies to accomplish context retrieval from an external data source. By default RAG uses SimilarityRetrieval that simply replicate the previous behaviour maintaining backward compatibility. But it depends now by its own interface so you can create custom implementation and inject it into the RAG.

Retrieval Documentation

New Features

Neuron CLI

V2 ships with practical developer experience improvements that address common friction points. The CLI tool brings the new “make” command that helps you scaffold common classes reducing boilerplate fatigue:

# Unix
php vendor/bin/neuron make:agent App\\Neuron\\MyAgent

# Windows
php .\vendor\bin\neuron make:agent App\\Neuron\\MyAgent

Evaluators

When building AI agents, evaluating their performance is crucial during this process. It's important to consider various qualitative and quantitative factors, including response quality, task completion, success, and inaccuracies or hallucinations.

Neuron introduces a system to create evaluators against test cases, so you can continues verify the output of your agentic entities overtime.

Evaluation Documentation

Structured Output validation rules

We introduced two new validation rules for structured output: WordsCount (works on string), and InRange (works on numeric).

Tool Max Tries

Agents now have a safety mechanism that tracks the number of times a tool is invoked during an execution session. If the agent exceeds this limit, execution is interrupted and an exception is thrown. By default the limit is 5 calls, and it count for each tool individually. You can customize this value with the toolMaxTries method.

try {

    $result = YouTubeAgent::make()
        ->toolMaxTries(5) // Max number of calls for each tool
        ->chat(...);
        
} catch (ToolMaxTriesException $exception) {
    // do something
}