ai-agentic-graph-rag

A space for learning GraphRAG with Semantic Kernel neo4j.

Aspire Parameters

Parameters in the Aspire appsettings.json file Parameters section follow a convention of SectionName-SettingName with the hyphen acting as a separator. The same pattern is used when passing the resources , with : used as the seperator.

AI settings

AI is configured to use one of several providers. This is based on https://chris-ayers.com/2025/07/06/aspire-with-lots-of-ai/ and the corresponding git repository https://github.com/codebytes/build-with-aspire.

See also AI integrations compatibility matrix.

To set this up:

• Add model integration nuget packages to the AppHost project.

  • Aspire.Hosting.GitHub.Models
  • Aspire.Hosting.Azure.AIFoundry
  • Aspire.Hosting.Azure.CognitiveServices
  • CommunityToolkit.Aspire.Hosting.Ollama

• Add appSettings..json files with the relevant settings for each provider. appsettings.Development.json can be used as the default.

  • appsettings.AzureOpenAI.json
  • appsettings.Development.json
  • appsettings.GitHubModels.json
  • appsettings.Ollama.json

• Change Properties\launchSettings.json to include profiles for each AI provider. Those profiles can be copied from the existing https profile, and the ASPNETCORE_ENVIRONMENT and DOTNET_ENVIRONMENT variables changed to match the on the appSettings files. This will allow the project to be run in Visual Studio using different AI providers by selecting the relevant profile.

• Add an extension class AIModelExtensions to the AppHost project. This will load the relevant AI model based on the configured provider.

• Modify Program.cs in the AppHost project to use the AIModelExtensions class to add the AI model to the service collection.

•   var aiService = builder.AddAIModel();
    var aiService = builder.AddAIEmbeddingModel();
    var (aiModel, aiEmbeddingModel) = builder.AddAIEmbeddingModel(); 
  

• For Github Models, make sure to add your GitHub API key either as a parameter in user secrets {name}-gh-apikey where name is the name of the deployment(s) in your configuration (e.g. chat-gh-apikey embedding-gh-apikey), or in the GITHUB_TOKEN environment variable.

• For Azure OpenAI, set the endpoint and ApiKey in user secrets (the ApiKey is optional; if not provided then the code will use the current user identity). Azure OpenAI also has optional settings AI:ModelVersion, AI:SkuName and AI:SkuCapacity that will be used if the models need to be provisioned.

• On the client side, add the following nuget packages:

  • Aspire.Azure.AI.OpenAI
  • Aspire.Azure.AI.Inference
  • CommunityToolkit.Aspire.Client.Ollama

Resilience

Timeouts can occur with Ollama models because generating responses or loading large models can take longer than the default 10-second resilience timeout in .NET Aspire or the 5-minute default model unloading time in Ollama. To address this a new policy extension AddOllamaResilienceHandler has been added and applied in AIServiceExtensions.AddAIProvider when Ollama is used.

See this issue.

For GitHub models, 429 (too many requests) responses can occur when rate limits are exceeded. EinsteinDataIngestionService has been extended to use a resilience pipeline when generating embeddings.

Graph database - Neo4j

How to get started with Neo4j see Build applications with Neo4j and .NET. For a suggested approach to dependency injection see Neo4j Data Access for Your .NET Core C# Microservice.

Neo4j can be run locally using Docker. See https://neo4j.com/docs/operations-manual/current/docker/introduction/

A dockerfile is included in the AppHost project and this will start Neo4j when the application starts.

To start Neo4j in docker from a command line use the following (backticks used for multi-line commands in Windows Terminal):

docker volume create neo4j_data `
  && docker volume create neo4j_logs `
  && docker volume create neo4j_import `
  && docker volume create neo4j_plugins

docker run -d --name neo4j-agentic-graphrag `
  -p 7474:7474 -p 7687:7687 `
  -v neo4j_data:/data `
  -v neo4j_logs:/logs `
  -v neo4j_import:/import `
  -v neo4j_plugins:/plugins `
  -e NEO4J_AUTH=neo4j/password `
  -e 'NEO4J_PLUGINS=["apoc", "apoc-extended", "graph-data-science"]' `
  -e apoc.import.file.enabled=true `
  -e apoc.import.file.use_neo4j_config=false `
  neo4j:2025.09.0

The Neo4j browser can be accessed at http://localhost:7474/browser/. Use the username neo4j and default password password.

Delete data by running this in the Neo4j data browser:

MATCH (n) DETACH DELETE n;
CALL apoc.schema.assert({},{},true) YIELD label, key RETURN *

Some of the code in this project was inspired by the book Essential GraphRAG. The code can be found at https://github.com/tomasonjo/kg-rag. Datasets for the author's earlier book on graphs can be found in https://github.com/tomasonjo/graphs-network-science.

Setting up Movies data

Movies data can be added by opening the ai-agentic-graph-rag\notebooks folder in vs code and running notebook Neo4j load movies.dib.

Note that the first step in the notebook copies files into the docker container; if the copy fails check that the container is running and the name of the container in the cp command is correct. The container name can be set in the appSettings parameters - for development this is "GraphDatabase-DockerContainerName": "neo4j-agentic-graphrag"; if no value is set the AppHost will create a default name.

Cypher

To set a parameter in Cypher in the neo4j browser:

:param name => 'Tom Hanks'

You can then use the parameter in a query:

MATCH (a:Person {name: $name})-[:ACTED_IN]->(m:Movie)
RETURN m.title AS movieTitle

Graph database - Memgraph

If using Memgraph as a replacement for Neo4j, see:

• https://stackoverflow.com/questions/74528361/how-can-i-connect-to-memgraph-from-my-c-sharp-application
• https://memgraph.com/blog/how-to-build-a-flight-network-analysis-graph-asp-net-application-with-memgraph-c-sharp-and-d3-js

Docker

See Add Dockerfiles for details on using Docker files with the AppHost project.

When the project was first run the dockerfile failed to build. A simple file was added to see if it would build and run; this file has been renamed to Dockerfile_simple_python and it loads and runs a simple python script main.py.

https://neo4j.com/docs/operations-manual/current/docker/introduction/

Code style and analysis

The ServiceDefaults project was created by the Aspire template and there are a number of code analysis warnings in the project. Those warnings have been suppressed in the project file.

Einstein sample

Inspired by the book Essential GraphRAG. Some code was adapted and translated from the book code repository:

• chapter 2
• chapter 3.

DevUI

Added DevUI based on Supercharging .NET Apps with DevUI and the New Microsoft Agent Framework.

TODO

• add more unit tests
• make the chunking IAsyncEnumerable based
  • link text
• Look at extracting and using UFO dataset
  • See chapter 5 of sql_book
  • SQL book to R
  • Import csv