Michigan Imaginary Energy Assistant

Prerequisites

This application requires an OpenAI API key to function properly. Ensure you have a valid API key before using the app.

Features

The Michigan Imaginary Energy Assistant provides a robust set of features designed to enhance data retrieval and ticket management. Key features include:

1. Hybrid Retrieval System

• Retrieves information from a vector database for efficient semantic search.
• Retrieves information from an SQLite database for structured query-based retrieval.
• Combines both sources to create a hybrid retrieval system, ensuring accurate and comprehensive responses.

2. Ticketing System

• Users can create new tickets for support or assistance.
• Users can retrieve the status of existing tickets, ensuring smooth issue tracking and resolution.

Usage

1. Set up the environment:

   • Create a Python environment using Conda:

     conda create --name <env_name> python=3.10.13
     conda activate <env_name>

   • Install dependencies:

     pip install -r requirements.txt

2. Seed the database (optional):

   • If you need to re-seed the database, navigate to the create_data directory and run:

     python createdb.py
     python loaddocs.py

3. Run the application:

   • Start the app using Streamlit:

     streamlit run app.py

   • The application will be available at http://localhost:8501

Project Structure

.
├── app.py
├── chroma_db
│   ├── ade9063b-75b9-45b4-9c88-c12d579e1a7e
│   │   ├── data_level0.bin
│   │   ├── header.bin
│   │   ├── length.bin
│   │   └── link_lists.bin
│   └── chroma.sqlite3
├── create_data
│   ├── createdb.py
│   ├── documents
│   │   ├── Energy_Efficiency_Tips.txt
│   │   ├── Green_energy_plan_FAQ.txt
│   │   └── Solar_Panel_Installation_Guide.pdf
│   └── loaddocs.py
├── energy_products.db
├── README.md
├── requirements.txt
└── testing
    └── main.py

5 directories, 15 files

Additional Details

API Key Configuration

• The application requires an OpenAI API key for execution.
• The API key can be entered in the sidebar input field.

Database Interaction

• Uses SQLite as the primary database.
• Supports querying multiple tables using the SQL agent.

Vector Search

• Utilizes ChromaDB for semantic search.
• Embedding model used: sentence-transformers/all-MiniLM-L6-v2.

Agent Architecture

• The implementation uses the ReAct agent from the LangChain framework.
• The agent dynamically decides when to use each tool based on the tool descriptions provided during definition.
• This approach allows the agent to efficiently retrieve information and generate responses based on the query context.
• The agent leverages a hybrid retrieval system, combining SQL queries for structured data and ChromaDB for unstructured data.
• The LangChain SQL agent executes queries against the database, while the vector retriever fetches semantically relevant documents.
• The system prioritizes structured data first and enhances it with unstructured document insights.

Retrieval Tool Details

• The retriever tool is responsible for extracting relevant information from both structured and unstructured data sources.
• For structured data, the multi_table(query) function leverages the LangChain SQL agent to execute SQL queries across multiple tables in SQLite.
• For unstructured data, retriever.get_relevant_documents(query) fetches semantically similar documents using ChromaDB.
• The results from both retrieval methods are combined to generate a holistic response to user queries.
• This dual approach ensures that responses are both factually accurate from structured data and contextually rich from unstructured knowledge sources.

Agent Decision Flow

• The diagram represents how user input is processed by the LLM.
• If the input requires a tool, the agent chooses between the retriever_tool and ticket_creation_tool.
• If no tool is needed, the response is generated directly.

Interacting with the Q&A System

• Users can submit natural language queries through the Streamlit UI.
• The system automatically determines whether to query the SQL database or retrieve vector-based knowledge.
• Example queries:
  • "I need assistance with solar panel installation"
  • "Retrieve all current tickets and all information"
  • "What are our current energy offers?"
  • "What is the difference between the Green Energy Plan and Fixed-Rate Plan?"

Ticket Creation Workflow

• Users can submit a request for support via the interface.
• The system logs the request in the energy_products.db database under the tickets table.
• Example:
  • User inputs: “I need assistance with solar panel installation.”
  • The system responds: “Ticket #5678 created successfully.”
• The ticket can be retrieved later using its ID.

Available Tools

• retriever_tool: Retrieves information from structured and unstructured sources.
• ticket_creation_tool: Allows users to create support tickets.

Query Execution

• Queries can retrieve information from both SQL and vector databases.
• Support tickets are stored in energy_products.db

Deliverables

The project source code is available in a GitHub repository or a zipped folder containing:

• Source code with implementation scripts.
• Database setup scripts for initializing and seeding the SQLite database.
• Example documents in the file system for vector search indexing.
• Instructions on environment setup, dependencies, and execution steps.
• Documentation (README) detailing:
  • How the RAG pipeline is built and operates.
  • Interaction steps for the Q&A system.
  • Ticket creation and tracking workflows.

For any additional troubleshooting or modifications, refer to the documentation or reach out to the development team.