Skip to content

dotnet

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History
 
 

dotnet

README.md

GraphRAG for .NET

A .NET 10 port of Microsoft GraphRAG — a graph-based retrieval-augmented generation (RAG) system that uses knowledge graphs to enhance LLM reasoning over complex datasets.

Solution Structure

Project Type Description
GraphRag Console App CLI entry point — orchestrates indexing, querying, and prompt-tuning workflows
GraphRag.Common Library Shared abstractions, configuration loading, hashing utilities, strategy discovery, and DI helpers
GraphRag.Cache Library Caching layer with memory, JSON file, and no-op implementations
GraphRag.Chunking Library Text chunking strategies (sentence-based, token-based) for document processing
GraphRag.Input Library Document ingestion — reads CSV, JSON, and plain text formats (core)
GraphRag.Llm Library LLM abstraction layer for completions, embeddings, tokenization, and templating
GraphRag.Storage Library Data persistence with file system and memory backends (core)
GraphRag.Vectors Library Vector store abstraction with filtering and search

Strategy Libraries (Plugin Assemblies)

Project NuGet Dependency Implements
GraphRag.Storage.AzureBlob Azure.Storage.Blobs IStorage
GraphRag.Storage.AzureCosmos Microsoft.Azure.Cosmos IStorage
GraphRag.Storage.Parquet Parquet.Net ITableProvider
GraphRag.Storage.Csv CsvHelper ITableProvider
GraphRag.Vectors.AzureAiSearch Azure.Search.Documents IVectorStore
GraphRag.Vectors.AzureCosmos Microsoft.Azure.Cosmos IVectorStore
GraphRag.Vectors.LanceDb (stub) IVectorStore
GraphRag.Llm.AzureOpenAi Azure.AI.OpenAI ILlmCompletion, ILlmEmbedding
GraphRag.Llm.SharpToken SharpToken ITokenizer
GraphRag.Llm.Scriban Scriban ITemplateEngine
GraphRag.Input.CsvHelper CsvHelper IInputReader
GraphRag.Input.Markdig Markdig IInputReader
GraphRag.Input.OpenXml DocumentFormat.OpenXml IInputReader
GraphRag.Nlp.Catalyst Catalyst INounPhraseExtractor
GraphRag.Graph.QuikGraph QuikGraph IGraphAlgorithms

Web Application

Project Type Description
GraphRag.SearchApp Blazor Server Interactive search UI — run Global, Local, DRIFT, and Basic RAG queries with community report explorer (Getting Started)

Prerequisites

Build

dotnet build

Test

# Run all tests
dotnet test

# Unit tests only
dotnet test tests/GraphRag.Tests.Unit

# Integration tests only
dotnet test tests/GraphRag.Tests.Integration

Run the CLI

dotnet run --project src/GraphRag -- <command> [options]

CLI Commands

Command Description
init Initialize a new GraphRAG project with default configuration
index Build a knowledge graph index from input documents
query Query the knowledge graph using natural language
prompt-tune Auto-tune prompts for your specific dataset

Examples

# Initialize a new project
dotnet run --project src/GraphRag -- init --root ./my-project

# Build an index
dotnet run --project src/GraphRag -- index --root ./my-project

# Query the graph
dotnet run --project src/GraphRag -- query --root ./my-project --query "What are the main themes?"

Run the Search App

The Search App provides a browser-based UI for querying and exploring indexed datasets.

cd src/GraphRag.SearchApp
dotnet run
# Open https://localhost:5001

See the full Search App Getting Started Guide for configuration, data preparation, and usage instructions.

Architecture

The solution follows a strategy pattern architecture where core libraries contain only interfaces and pure-.NET implementations, while all 3rd-party integrations (Azure SDKs, NLP libraries, etc.) are isolated into separate plugin assemblies:

  • Core libraries depend only on abstractions — no 3rd-party NuGet packages
  • Strategy libraries each encapsulate one external dependency behind a core interface
  • Runtime discovery auto-scans GraphRag.*.dll assemblies for [StrategyImplementation]-decorated classes
  • Configuration maps strategy keys to implementations in settings.yaml

Pipeline

  • Input → Documents are read and normalized via IInputReader
  • Chunking → Text is split into manageable chunks for processing
  • LLM → Completions, embeddings, and tokenization via ILlmCompletion/ILlmEmbedding
  • Storage → Indexed graphs and documents are persisted via IStorage
  • Vectors → Embeddings are stored and searched via IVectorStore
  • Cache → Intermediate results are cached to reduce redundant API calls

All configuration is driven by YAML/JSON settings files created during init.

License

MIT