Graph Chat

This hack explores:

• Populating vertices and edges in a Graph database from an event source
• Using that Graph to power a more personalised LLM assistant, using Microsoft Agent Framework
• Extracting preferences from the user chat and persisting them back to the graph for future usage.

It makes use of the anonymised H&M dataset, available on Kaggle.

Agents

• user_agent:

  • Looks up user purchses from the graph
  • Recommends items based on customer segment queries
  • Finds similar items to existing ones
  • Calls preference extraction as an agent inside a memory layer

• dynamic_query_agent:

  • Experiment to test handing the agent the graph schema and a user query, and dynamically generating a gremlin query. Not safe, there be dragons here :)

• signals_extraction_agent:

  • Pass the conversation to this agent along with a model schema for preferences, and have it extract those customer preferences from the chat.

High Level Architecture

Graph Structure

Usage

In the super handy Dev UI extension for Agent Framework:

Directory Structure

├── infra/                 # Terraform infrastructure definitions
├── src/
│   ├── agents/            # Agents, tools & memory
│   └── ingestion/         # Data pipeline
│       ├── ingest/        # CSV → Event Hub producer
│       ├── consume/       # Event Hub consumer → Cosmos DB
│       └── models/        # Event data models
└── tests/                 # Unit, integration and ai evals

Getting Started

1. Clone the repository

   git clone https://github.com/damoodamoo/graph_chat.git
   cd graph_chat

2. Open in dev container

   • Open the project in VS Code
   • When prompted, click "Reopen in Container"
   • Or use Command Palette: Dev Containers: Reopen in Container

3. Authenticate with Azure

   az login

4. Configure environment variables

   • Create a .env file in the project root:

   cp .env.sample .env
   

5. Deploy infrastructure

   task infra:deploy

   Deploying the infra creates the app.env file that the ingestion and chat agents need.

6. Download and reduce the source data from Kaggle

   • Get your Kaggle API key and set it in .env

   • Download the dataset

   task data:download

   Now you have the full (and reduced) dataset downloaded.

7. Run the 'ingest' and 'consume' flows to populate the graph. In one terminal session, start the consumer...

   task consume:all

   ...and in another session, start the ingest (producer):

   task ingest:all

8. Run the agent

   task agent:dev
   

   ...ask it some stuff.

9. Run the evals Currently there is an initial set of evals for preference extraction.

   task eval
   

   Other unit / integration tests are run via task test.

---

This is an experimental project and is intended for upskilling.