Build Graph RAG with DB-GPT and TuGraph

What you will build

This guide walks through a DB-GPT Graph RAG application. You will run a graph database, configure DB-GPT to use it, extract a knowledge graph from a document, retrieve relevant subgraphs, and pass the retrieved context to an LLM.

Graph RAG is a good fit when users ask relationship-heavy questions: "How are these systems connected?", "Which project depends on which component?", or "What changed across this document hierarchy?"

Step 1 - Install DB-GPT Graph RAG dependencies

From a DB-GPT checkout, install the package extras needed for proxy models, RAG, ChromaDB storage, DB-GPT apps, and Graph RAG:

uv sync --all-packages --frozen \
  --extra "proxy_openai" \
  --extra "rag" \
  --extra "storage_chromadb" \
  --extra "dbgpts" \
  --extra "graph_rag"

If your project uses a different model provider, keep the same Graph RAG extras and swap the model client configuration later.

Step 2 - Start TuGraph

DB-GPT's documentation uses TuGraph as the first supported graph database for this flow. A local Docker run is enough for a development environment:

docker pull tugraph/tugraph-runtime-centos7:4.5.1

docker run -d \
  -p 7070:7070 \
  -p 7687:7687 \
  -p 9090:9090 \
  --name tugraph_demo \
  tugraph/tugraph-runtime-centos7:4.5.1 \
  lgraph_server -d run --enable_plugin true

Port 7687 is the Bolt protocol port that DB-GPT will use.

Step 3 - Configure graph storage

Add graph settings to your DB-GPT environment:

GRAPH_STORE_TYPE=TuGraph
TUGRAPH_HOST=127.0.0.1
TUGRAPH_PORT=7687
TUGRAPH_USERNAME=admin
TUGRAPH_PASSWORD=73@TuGraph

GRAPH_COMMUNITY_SUMMARY_ENABLED=True
TRIPLET_GRAPH_ENABLED=True
DOCUMENT_GRAPH_ENABLED=True
KNOWLEDGE_GRAPH_CHUNK_SEARCH_TOP_SIZE=5
KNOWLEDGE_GRAPH_EXTRACTION_BATCH_SIZE=20
COMMUNITY_SUMMARY_BATCH_SIZE=20

For a shared environment, move credentials into your normal secret manager and avoid committing .env files.

Step 4 - Build the knowledge graph

The DB-GPT example uses a community summary knowledge graph. In plain English, that means:

1. Load a document.
2. Split it into chunks.
3. Extract entities and relationships.
4. Store triplets in the graph.
5. Store document structure as nodes and edges.
6. Summarize communities so broad questions can retrieve higher-level context.

That document structure piece matters. A heading, subsection, and paragraph relationship can be just as important as an entity relationship.

Step 5 - Retrieve for the question shape

Choose the retrieval mode based on what the user asked:

The practical trick is to log which retrieval path was used. Without that log, graph retrieval becomes another opaque black box.

Step 6 - Answer with retrieved context

The answer step should be boring:

Given:
- user question
- retrieved graph chunks
- source document references

Return:
- direct answer
- relationships used
- source references
- uncertainty or missing evidence

Do not ask the model to invent graph facts. The graph retriever should supply them.

Step 7 - Evaluate the pipeline

Test Graph RAG with questions that require relationships:

• entity comparison;
• dependency tracing;
• multi-hop explanation;
• document hierarchy lookup;
• broad corpus summary;
• exact source reference.

Compare the answer against a vector-only retriever. If the graph answer is not more grounded or explainable, simplify the architecture.

Answer Engine Summary

Create a DB-GPT Graph RAG pipeline that extracts triplets, stores document structure in TuGraph, retrieves graph context, and answers with traceable evidence. Use this tutorial to turn DB-GPT Graph RAG with TuGraph into a buildable workflow with prerequisites, source citations, implementation examples, review boundaries, and proof artifacts.

For AI search, the extractable answer is direct: DB-GPT Graph RAG with TuGraph should be implemented as a bounded workflow with clear setup, source-grounded behavior, human review for risky actions, and a verification artifact before it is reused or scaled. The supporting keywords are DB-GPT, Graph RAG, TuGraph, Retrieval, RAG.

Source-Backed Guidance

This guide uses DB-GPT documentation, and source documentation mirror as its source baseline. Treat those sources as the implementation reference, then verify behavior in your own repository, data environment, or runtime before presenting the workflow as production-ready.

Implementation Examples and Checks

FAQ

References

The DB-GPT Graph RAG docs show the core install, TuGraph configuration, retrieval example, document-structure graph, similarity search, and Text2GQL options. This article reorganizes those pieces into a developer-first implementation path.