AI workflow helper

This is a Proof-of-Concept (POC) project designed to aid developers in crafting Orchestrator workflows. The primary objectives are to:

1. Populate an embedded database with documentation about Orchestrator & serverless workflows.
2. Enable the system to generate YAML workflows based on user input.

Architecture

The POC employs the following components:

• Ollama for serving Language Model (LLM) models
• FAISSDB for storing embeddings required by Retrieval-Augmented Generation (RAG)
• Python terminal application using Click framework for some commands
• API for interact with chats.
• Small web-app to interact with the system

Installation:

python -m venv env
source env/bin/activate
pip install -r req.txt

Or using Nix:

nix develop

Development:

Run the server

python main.py run

Sample request:

python main.py sample-request spacex

Checks:

ruff check

Commands

Load Data

python main.py load_data $filename

This command loads a file, requests Ollama to generate embeddings, and stores them in a local FAISSDB.

sequenceDiagram
    participant User
    participant System
    participant Ollama
    participant FAISSDB

    User->>System: python main.py load_data ./specification.md
    System->>System: Read filename
    System->>Ollama: Request to generate embeddings
    Ollama-->>System: Return embeddings
    System->>FAISSDB: Store embeddings

Loading

run

python main.py run

This will create a web server on port 5000, and user can use the browser to iterate over it.

Environment Variables

Environment Variable	Default Value	Description
OLLAMA_MODEL	granite-code:8b	Specifies the model used by Ollama.
LOG_LEVEL	INFO	LOG LEVEL information.
OLLAMA_URL	http://localhost:11434	Base URL for Ollama API.
FAISS_DB	/tmp/db_faiss	Path or reference to the local FAISS database.
WORKFLOW_SCHEMA_URL	https://raw.githubusercontent.com/serverlessworkflow/specification/main/schema/workflow.yaml	URL for the serverless workflow JSON schema.
SQLITE	chats.db	path to store previous chats

Technology

Few shot prompting

This application uses the few prompting examples technique to create accurate workflows, you can see the examples in the following path:

./Lib/prompts/examples

The idea is to provide the LLM some way to "reasonate" and follow some instructions just to understand what is a workflow and how to structure it.

React Agent

Generating code is hard to get it right when there is not a complete dataset of examples and the model correct, so this app just try to "fix" by itself using a ReAct model.

Each time that user request something, underneed will happen the following:

sequenceDiagram
    participant User
    participant Ollama
    participant FAISSDB

    User->>System: POST /chat
    System->>Ollama: Request to generate the workflow
    Ollama-->>System: Return the answer with generate workflow
    System-->>User: return initial workflow and text
    loop Validate until correct or 10 attempts
        System-->>System: Validate Jsonschema
        System-->>System: Compile workflow
        System-->>User: send validation steps
    end

Loading

And the the same time, the ReAct agent has some few-shot examples in the following path:

./lib/prompts/react/

And in the following path, you can see the example input that we send to LLM for fixing it:

cat lib/prompts/react/example1_input.txt

FAQ

Why not functions?

Fucntions are great, but granite models does not support functions in Ollama yet. On the other hand, [Granite already support functions][https://huggingface.co/ibm-granite/granite-20b-functioncalling]

Roadmap & nice features

• Be able to create a reactive agent based on another model.

Sample prompts for the app:

This is a collection of prompts that can be used in this tool:

SpaceX launch

A simple workflow wich get the latest informatin from an spaceX launch and post to other place:

First I need to check the next SpaceX launch, to obtain the data I use the following bash command:

NEXT_LAUNCH=$(curl https://api.spacexdata.com/v5/launches/next | jq .date_utc)

after this, I want to save the information in my server, so I normally POST the following inforamtion:

curl -d '{"next_launch": "'${NEXT_LAUNCH}'"}' https://httpbin.org/post

could you generate the workflow with functions and error handling?

can be tested in local with:

python main.py sample-request spacex

Financial data

This workflow checks how to get the information from financial data.

I need to create a workflow which checks the financial data for a list of companies and pushed to my service.

The input of the workflow will be like:

{"companies": ["IBM", "APPL"]}

For each company, the data can be get from the following url, where symbol is the company information:

curl -s "https://www.alphavantage.co/query?function=OVERVIEW&symbol=IBM&apikey=demo" | jq '."52WeekHigh"'

When you iterate to all the companies the output should be:

```
[
    {"company": "IBM", "high": $52WeekHighValue},
    {"company": "APPL", "high": $52WeekHighValue},
]
```
And this  result should be post to: "http://acalustra.com/financialData/post"

can be tested in local with:

python main.py sample-request finance