Starting workflow from AI assistant chat

[mpoweredo]

First of all, thank you very much for this awasome library!

Currently im making app for creating project estimates and now im working on adding AI assistant chat that can do tools calls, where user can prompt and tweak things inside estimate. So far so good I didnt have any problems.

In my app there is dedicated module for generating estimate using AI. When you provide docs to generate estimate for, user have to select bunch of options to start generating estimate and I want to move this whole dedicated module to AI assistant chat but kinda I do not know how to do this properly.

I want to move this dedicated module to AI assistant chat and want this flow:

STEP #1: User provide .docx files to the chat with prompt "generate estimate from this .docx"
STEP #2: I need to analyze prompt provided by user and if:

[FALSE CONDITION] prompt does not have enough information about what this estimate should look like, AI assistant need to ask user follow up question with answers that I need, for example "Provide granularity of project, should this be in mandays or hours, what roles are working in this project, do you have any additional instructions", etc.
[TRUE CONDITION] If user provide all this information in prompt, we can go to another step

STEP #3: Now when I have all information needed, I can generate estimate

I was reading neuron docs and there is stuff like "Workflows" and "Human In The Loop". I think it fits well for my scenario because I have sets of tasks to do to achive one thing (workflow), there is conditional logic based on the user input (Conditional Edges) and human followup questions (Human in the loop) - Should I use workflow here? What would you recommend?

I had a plan to add two tool calls to my AI assistant.

1. StartGenerateEstimateWorkflowTool - Tool for starting workflow, when user provide file with prompt "generate estimate for this .docx". This tool would start the GenerateEstimateWorkflow.
2. ResumeGenerateEstimateWorkflowTool - Tool for resuming workflow, it takes workflowId as input.

GenerateEstimateWorkflow would have nodes like CheckUserPrompt and based on the user prompt it would interrupt workflow or continue.

What do you recommend here? Is my approach good? Is there better way to achive that what I need?

Thank you for good library one more time!

[ilvalerione]

Hi, thank you for your feedback, and for posting your use case! Very interesting.

Important news - NeuronAI V2 is coming

Yes, Workflow is designed to help you implement these higly customized logic. At this point there are no solutions in PHP to solve this scenario. But workflow in Neuron V1 was an experiemntal component.

We are about to release Neuron V2, and it finally includes the stable architecture of the Neuron Workflow. It's easier to work with compared to V1, and it's incredibly powerful.

Here is the link to V2 documentation to better understand the changes and start learning how to create your workflow.

https://docs.neuron-ai.dev/v2/overview/readme/upgrade-guide

To start working with V2 you need to add the following configuration to your composer file:

{
	...
	
	"repositories": [
		{
			"type": "git",
			"url": "https://github.com/inspector-apm/neuron-ai"
		},
	]
	
	...
}

Then use the following entry in the require section:

"inspector-apm/neuron-ai": "2.*@dev",

then composer update.

I strongly recommend to start your MVP with V2 and when the final release is generaly available you can change your composer file to use the official version.

It will be released soon, so stay connected with us to get the news:

• Newsletter: https://neuron-ai.dev
• Linkedin: https://www.linkedin.com/company/neuron-ai-php-framework
• X: https://x.com/neuronai_php
• Instagram: https://www.instagram.com/neuronai_php_adk/

About Your Use Case

Reading about your initial idea I understand you want to have an Agent in front of the process, which is responsible to execute or resume the workflow, thanks to the two tools you mention.

These approach can help to have the "front" AI agent automatically manage the context of the conversation with a user. But it's less flexible and it makes impossible to customize the final report composition since the result of the workflow is always returned to the "front" agent as tool call result. Then the agent will apply it's judge to provide the final response.

Workflow should manage the entire process. You have to think to the workflow not just as a component to execute nodes. Workflow is the agent, but instead of a single LLM interface, you can create an entity that manage multiple interactions with LLMs for different purposes.

It's like a component that help you create a brain specialized in a task (creating estimates in your case).

I know this is a more complex way to solve the problem, but it's the way to go I believe.

When a user ask for an estimate you should start a workflow providing an initial state in input, with the file, and the user query.

The nodes could be:

1. CheckRequirementsNode
   Check if the user provided all the information for a complete estimate. If yes move forward to the estimate generation. If no, interrupt and ask for the missing information.

2. Planning
   Creates the structure of the report

3. Content Generation
   Generates content for each section using provided files and information

4. Format
   Compile the final report

Here is an example of a Deep Research Agent built with Neuron Workflow V2: https://github.com/inspector-apm/deep-research-agent

I hope it could be helpful. Anyway let us know your questions, or feel free to post here the progress of your implementation.

[mpoweredo]

thanks for the detailed response, im gonna check the new version of neuron