Overview
04/05/2026
GraphRAG version 1.2.0
This guide provides architectural and technical information about implementing GraphRAG workflows.
GraphRAG combines knowledge graph technology with Large Language Models (LLMs) to deliver explainable and graph-grounded answers to user questions. The workflows described in this guide orchestrate this interaction by controlling how queries are interpreted, how graph context is retrieved, and how LLM calls are made and customized.
A workflow in GraphRAG is the primary element of automated business logic. The GraphRAG product is built on several dozen workflows, each encapsulating specific logic and automation.
Because it is impossible to provide a one-size-fits-all solution, the GraphRAG implementation is based on a suggested base or default solution. This solution should be modified and adjusted to meet different project needs.
To understand how and where to easily and effectively modify the base workflows, you need to understand the main principles and ideas behind the base solution.
See the Configuration Guide for information on installation, credentials, service setup, and all configurable properties.
When you first sign in to GraphRAG Workflows, you will see the Workflows tab and its list of available workflows, along with other tabs that provide additional information and configuration options. Across the top of the screen are various statistics about recent usage of the workflows. In the lower- and upper-left are icons that offer additional features; tooltips for each tell you what they do.
 |
You can create a new workflow with the orange Create workflow button in the upper right and you can click any listed workflow to show it on the main screen where you can edit and execute it. This guide explains how you can use and customize the workflows provided with GraphRAG to combine your knowledge graphs with LLMs to build a system that provides grounded answers to questions submitted to it.
GraphRAG was never meant to be a black box. It was designed and built with these goals in mind:
Full transparency: Every action and decision is visible and traceable. All information is provided to the end user. Every answer can be explained and traced back.
Ease of customization: You can turn steps on and off, add additional steps, and inject custom information. The basic implementation is a starting point, not a finished product.
 |
These four stages represent the default configuration, which is customizable.
Stage | Purpose |
|---|---|
1. Pre-processing | Decide whether to proceed. Runs input guardrails; may include additional checks and syncs. |
2. Data Gathering & Context Engineering | Collect information from all sources: taxonomies, vector indexes, knowledge graphs, various types of databases, conversation and chat history. Can be extended to include other sources of data. |
3. Main Answer | At this stage all collected data is prepared for processing and the answer is generated. |
4. Post-Answer Processing | Analyze the answer (output guardrails), extract concepts from the final answer, and generate follow-up questions. |
To keep the Main workflow manageable, each stage delegates to sub-orchestrator workflows.