Introduction to DB-GPT

Relevant source files

DB-GPT is an open-source AI-native data application development framework designed to build infrastructure for large language model applications. This document provides a high-level overview of the system's architecture, core capabilities, and design philosophy.

Scope: This page covers the fundamental concepts, architecture, and capabilities of DB-GPT. For detailed installation instructions, see Installation. For deep dives into specific subsystems like RAG, AWEL, or Multi-Agents, refer to their respective sections (RAG, AWEL and Core Framework, Multi-Agents).

What is DB-GPT?

DB-GPT is a comprehensive framework for developing data-driven AI applications with minimal code. Built as a modular monorepo, it provides production-ready components for:

Model Management: Unified interface for 50+ LLMs (local and API-based)
Knowledge Processing: RAG (Retrieval-Augmented Generation) with multiple storage backends
Data Intelligence: GBI (Generative Business Intelligence) with Text2SQL capabilities
Workflow Orchestration: AWEL (Agentic Workflow Expression Language) for declarative workflows
Agent Framework: Data-driven multi-agent collaboration system

The framework enables enterprises and developers to build bespoke applications in the "Data 3.0" era, combining large models with structured and unstructured data sources.

Sources: README.md53-59 README.md68-81

System Architecture Overview

DB-GPT follows a layered architecture organized as a monorepo with seven specialized packages.

High-Level Architecture

Architecture Principles:

Modularity: Each package has a well-defined responsibility and can be installed independently
Abstraction: Core interfaces in dbgpt-core allow pluggable implementations in dbgpt-ext
Service-Oriented: Business logic separated into services in dbgpt-serve
Deployment Flexibility: Support for Docker, source code, and PyPI installation

Sources: README.md62-66 Diagram 1 from provided context

Monorepo Package Structure

DB-GPT v0.7.0 introduced a monorepo structure with seven packages, managed using the uv package manager for faster dependency resolution.

Package Hierarchy

Package	Purpose	Key Components
`packages/dbgpt-core`	Foundational abstractions and interfaces	AWEL engine, Agent framework, Storage interfaces (`StorageInterface`, `SQLAlchemyStorage`)
`packages/dbgpt-ext`	Pluggable storage and connector implementations	Vector stores (Milvus, Chroma, etc.), Knowledge graphs (TuGraph, Neo4j), Data sources (MySQL, PostgreSQL, etc.)
`packages/dbgpt-serve`	Service layer for business logic	RAG service, Evaluation service, Storage manager
`packages/dbgpt-app`	Main application entry point	`dbgpt_server.py`, Knowledge service, Application routing
`packages/dbgpt-client`	Client SDK for API access	Python client for REST API
`packages/dbgpt-accelerator`	Hardware acceleration modules	vLLM, Flash Attention, Quantization (BitsAndBytes, GPTQ)
`packages/dbgpt-sandbox`	Isolated execution environment	Code execution sandbox

Installation and Dependency Management

DB-GPT uses uv for fast dependency management with optional extras. Installation follows the pattern:

Configuration is managed through TOML files in the configs/ directory:

configs/dbgpt-proxy-openai.toml - OpenAI proxy model configuration
configs/dbgpt-proxy-deepseek.toml - DeepSeek proxy configuration
configs/dbgpt-local-glm.toml - Local GLM model configuration

Sources: README.md117-128 docs/docs/quickstart.md29-30 docs/docs/quickstart.md77-95

Core Capabilities

RAG (Retrieval-Augmented Generation)

DB-GPT provides a complete RAG framework for building knowledge-based applications. The pipeline supports:

Document Ingestion: Multiple file formats (.pdf, .md, .docx, .csv, .html) via KnowledgeFactory
Text Processing: ChunkManager for intelligent text splitting
Embedding: Integration with OpenAI, HuggingFace, and local embedding models
Storage: Multiple vector stores (Milvus, Chroma, Elasticsearch, PGVector, Weaviate, OceanBase) via IndexStoreBase
Retrieval: Hybrid strategies including semantic search, time-weighted retrieval, graph-based retrieval, and BM25 full-text search
Generation: Context-enhanced LLM responses through EmbeddingAssembler

For detailed RAG implementation, see RAG Pipeline and Knowledge Management.

Sources: README.md70 Diagram 2 from provided context

GBI (Generative Business Intelligence)

The GBI system bridges natural language queries with structured data sources through an intelligent Text2SQL pipeline:

Query Understanding: Natural language parsing and intent detection
Schema Linking: Automatic mapping of entities to database tables/columns using metadata
SQL Generation: LLM-powered generation achieving 82.5% accuracy on Spider benchmark
Query Execution: Connection pooling with error handling and retry logic
Visualization: Chart generation using GPT-Vis protocol with interactive dashboards

Supported data sources include MySQL, PostgreSQL, Oracle, MSSQL, ClickHouse, DuckDB, Hive, and Spark.

For Text2SQL implementation details, see Generative Business Intelligence (GBI).

Sources: README.md72-74 README.md168-171 Diagram 6 from provided context

Multi-Agent Framework

DB-GPT implements a data-driven multi-agent system with:

Agent Profiles: Role definitions with capabilities and constraints
Agent Memory: Short-term context and long-term knowledge storage
Agent Tools: Function calling for external integrations
Planning & Reasoning: Multi-step task decomposition
Collaboration: Agent orchestration via communication bus and shared state

Agent types include Data Agents, Plugin Agents, Code Agents, Chat Agents, and Custom Agents.

For multi-agent implementation, see Multi-Agents and AWEL Workflows.

Sources: README.md76 README.md172-174 Diagram 4 from provided context

AWEL (Agentic Workflow Expression Language)

AWEL is a declarative workflow language for composing complex data processing pipelines as Directed Acyclic Graphs (DAGs):

DAG Builder: Visual workflow definition
Operators: Transform, Join, Branch, Streamify, Unstreamify operators
Triggers: HTTP, scheduled, and event-based triggers
Execution: Runtime DAG executor with lifecycle management

AWEL enables developers to compose sophisticated workflows that integrate RAG, agents, models, and data sources.

For AWEL details, see Core Framework and AWEL.

Sources: README.md55 README.md200-203 Diagram 4 from provided context

SMMF (Service-oriented Multi-model Management Framework)

SMMF provides unified model management supporting 50+ LLMs through two deployment strategies:

Local Deployment:

HuggingFace Transformers
vLLM (high-throughput inference)
LLAMA.cpp (CPU/Metal inference)
MLX (Apple Silicon optimization)

API Proxies:

OpenAI, DeepSeek, Qwen, Ollama
Chinese providers: Baidu (文心), Alibaba (通义), Zhipu (智谱), Xunfei (讯飞)

The framework includes hardware acceleration through Flash Attention, quantization (BitsAndBytes 8-bit/4-bit, GPTQ), and CUDA support (11.8, 12.1, 12.4).

For model configuration, see Service-oriented Multi-model Management Framework (SMMF).

Sources: README.md180-292 README.md-zh.md191-318 Diagram 3 from provided context

Technology Stack and Dependencies

Core Technologies

Key Dependencies by Package

dbgpt-core: sqlalchemy, pydantic, fastapi
dbgpt-ext: Optional extras for storage (storage_milvus, storage_chromadb, graph_rag)
dbgpt-accelerator: vllm, flash-attn, bitsandbytes, gptq
dbgpt-app: fastapi, uvicorn, sqlalchemy

The default installation uses SQLite for metadata storage (no external database required). Production deployments can switch to MySQL or PostgreSQL via configuration.

Sources: packages/dbgpt-core/src/dbgpt/storage/metadata/db_storage.py1-49 docs/docs/quickstart.md27-28 Diagram 5 from provided context

Entry Points and Configuration

Main Application Entry

The primary entry point is dbgpt_server.py:

packages/dbgpt-app/src/dbgpt_app/dbgpt_server.py

Start the server using:

Or directly:

Sources: docs/docs/quickstart.md137-147 docs/docs/installation/sourcecode.md106-116

Configuration Files

Configuration follows a TOML-based system in the configs/ directory:

Configuration File	Purpose	Key Sections
`configs/dbgpt-proxy-openai.toml`	OpenAI proxy setup	`[models.llms]`, `[models.embeddings]` with `api_key`
`configs/dbgpt-proxy-deepseek.toml`	DeepSeek proxy setup	LLM and embedding configuration
`configs/dbgpt-local-glm.toml`	Local GLM-4 model	HuggingFace model paths with `provider = "hf"`

Example configuration structure:

For storage configuration:

Sources: docs/docs/quickstart.md124-135 docs/docs/installation/integrations/milvus_rag_install.md27-37 docs/docs/installation/integrations/oceanbase_rag_install.md27-37

Storage Architecture

DB-GPT implements a flexible storage abstraction through StorageInterface in dbgpt-core:

The SQLAlchemyStorage class provides the core metadata persistence:

This abstraction allows seamless switching between storage backends through configuration without code changes.

For detailed storage implementations, see Storage Architecture and Databases.

Sources: packages/dbgpt-core/src/dbgpt/storage/metadata/db_storage.py21-49 Diagram 2 from provided context

Deployment Options

DB-GPT supports multiple deployment modes:

1. Source Code Deployment

Sources: docs/docs/installation/sourcecode.md14-21 docs/docs/installation/sourcecode.md69-89

2. Docker Deployment

Multi-stage Dockerfiles support different installation modes:

Base mode: Core dependencies only
OpenAI mode: Proxy model support
vLLM mode: High-throughput local inference
LLAMA.cpp mode: CPU/Metal inference
Full mode: All features

Build and run:

Sources: Diagram 5 from provided context, docs/sidebars.js127-140

3. Docker Compose

Orchestrates multiple services (database, webserver) with volume management:

The docker-compose.yml file defines service dependencies and networking.

Sources: docs/sidebars.js131-134

4. PyPI Installation

For production deployments, install via pip:

Individual packages can be installed separately:

Sources: Diagram 5 from provided context

For comprehensive deployment instructions, see Deployment and Configuration.

Getting Started

To begin using DB-GPT:

Install Dependencies: Follow the Source Code Installation guide or use Docker
Configure Model: Choose between proxy models (OpenAI, DeepSeek) or local models (GLM-4, etc.) in configuration files
Start Webserver: Run uv run dbgpt start webserver --config <config-file>
Access UI: Visit http://localhost:5670 in your browser

For detailed quickstart instructions, see Quickstart.

For development setup and contributing, see Development Guide.

Sources: docs/docs/quickstart.md1-217 README.md139-159

Additional Resources

GitHub Repository: https://github.com/eosphoros-ai/DB-GPT
Documentation: http://docs.dbgpt.cn/docs/overview/
Paper: https://arxiv.org/pdf/2312.17449.pdf
Community: https://github.com/eosphoros-ai/community

Sources: README.md49 README.md323-349

Introduction to DB-GPT

Relevant source files

What is DB-GPT?

DB-GPT is a comprehensive framework for developing data-driven AI applications with minimal code. Built as a modular monorepo, it provides production-ready components for:

Model Management: Unified interface for 50+ LLMs (local and API-based)
Knowledge Processing: RAG (Retrieval-Augmented Generation) with multiple storage backends
Data Intelligence: GBI (Generative Business Intelligence) with Text2SQL capabilities
Workflow Orchestration: AWEL (Agentic Workflow Expression Language) for declarative workflows
Agent Framework: Data-driven multi-agent collaboration system

The framework enables enterprises and developers to build bespoke applications in the "Data 3.0" era, combining large models with structured and unstructured data sources.

Sources: README.md53-59 README.md68-81

System Architecture Overview

DB-GPT follows a layered architecture organized as a monorepo with seven specialized packages.

High-Level Architecture

Architecture Principles:

Modularity: Each package has a well-defined responsibility and can be installed independently
Abstraction: Core interfaces in dbgpt-core allow pluggable implementations in dbgpt-ext
Service-Oriented: Business logic separated into services in dbgpt-serve
Deployment Flexibility: Support for Docker, source code, and PyPI installation

Sources: README.md62-66 Diagram 1 from provided context

Monorepo Package Structure

DB-GPT v0.7.0 introduced a monorepo structure with seven packages, managed using the uv package manager for faster dependency resolution.

Package Hierarchy

Package	Purpose	Key Components
`packages/dbgpt-core`	Foundational abstractions and interfaces	AWEL engine, Agent framework, Storage interfaces (`StorageInterface`, `SQLAlchemyStorage`)
`packages/dbgpt-ext`	Pluggable storage and connector implementations	Vector stores (Milvus, Chroma, etc.), Knowledge graphs (TuGraph, Neo4j), Data sources (MySQL, PostgreSQL, etc.)
`packages/dbgpt-serve`	Service layer for business logic	RAG service, Evaluation service, Storage manager
`packages/dbgpt-app`	Main application entry point	`dbgpt_server.py`, Knowledge service, Application routing
`packages/dbgpt-client`	Client SDK for API access	Python client for REST API
`packages/dbgpt-accelerator`	Hardware acceleration modules	vLLM, Flash Attention, Quantization (BitsAndBytes, GPTQ)
`packages/dbgpt-sandbox`	Isolated execution environment	Code execution sandbox

Installation and Dependency Management

DB-GPT uses uv for fast dependency management with optional extras. Installation follows the pattern:

Configuration is managed through TOML files in the configs/ directory:

configs/dbgpt-proxy-openai.toml - OpenAI proxy model configuration
configs/dbgpt-proxy-deepseek.toml - DeepSeek proxy configuration
configs/dbgpt-local-glm.toml - Local GLM model configuration

Sources: README.md117-128 docs/docs/quickstart.md29-30 docs/docs/quickstart.md77-95

Core Capabilities

RAG (Retrieval-Augmented Generation)

DB-GPT provides a complete RAG framework for building knowledge-based applications. The pipeline supports:

Document Ingestion: Multiple file formats (.pdf, .md, .docx, .csv, .html) via KnowledgeFactory
Text Processing: ChunkManager for intelligent text splitting
Embedding: Integration with OpenAI, HuggingFace, and local embedding models
Storage: Multiple vector stores (Milvus, Chroma, Elasticsearch, PGVector, Weaviate, OceanBase) via IndexStoreBase
Retrieval: Hybrid strategies including semantic search, time-weighted retrieval, graph-based retrieval, and BM25 full-text search
Generation: Context-enhanced LLM responses through EmbeddingAssembler

For detailed RAG implementation, see RAG Pipeline and Knowledge Management.

Sources: README.md70 Diagram 2 from provided context

GBI (Generative Business Intelligence)

The GBI system bridges natural language queries with structured data sources through an intelligent Text2SQL pipeline:

Query Understanding: Natural language parsing and intent detection
Schema Linking: Automatic mapping of entities to database tables/columns using metadata
SQL Generation: LLM-powered generation achieving 82.5% accuracy on Spider benchmark
Query Execution: Connection pooling with error handling and retry logic
Visualization: Chart generation using GPT-Vis protocol with interactive dashboards

Supported data sources include MySQL, PostgreSQL, Oracle, MSSQL, ClickHouse, DuckDB, Hive, and Spark.

For Text2SQL implementation details, see Generative Business Intelligence (GBI).

Sources: README.md72-74 README.md168-171 Diagram 6 from provided context

Multi-Agent Framework

DB-GPT implements a data-driven multi-agent system with:

Agent Profiles: Role definitions with capabilities and constraints
Agent Memory: Short-term context and long-term knowledge storage
Agent Tools: Function calling for external integrations
Planning & Reasoning: Multi-step task decomposition
Collaboration: Agent orchestration via communication bus and shared state

Agent types include Data Agents, Plugin Agents, Code Agents, Chat Agents, and Custom Agents.

For multi-agent implementation, see Multi-Agents and AWEL Workflows.

Sources: README.md76 README.md172-174 Diagram 4 from provided context

AWEL (Agentic Workflow Expression Language)

AWEL is a declarative workflow language for composing complex data processing pipelines as Directed Acyclic Graphs (DAGs):

DAG Builder: Visual workflow definition
Operators: Transform, Join, Branch, Streamify, Unstreamify operators
Triggers: HTTP, scheduled, and event-based triggers
Execution: Runtime DAG executor with lifecycle management

AWEL enables developers to compose sophisticated workflows that integrate RAG, agents, models, and data sources.

For AWEL details, see Core Framework and AWEL.

Sources: README.md55 README.md200-203 Diagram 4 from provided context

SMMF (Service-oriented Multi-model Management Framework)

SMMF provides unified model management supporting 50+ LLMs through two deployment strategies:

Local Deployment:

HuggingFace Transformers
vLLM (high-throughput inference)
LLAMA.cpp (CPU/Metal inference)
MLX (Apple Silicon optimization)

API Proxies:

OpenAI, DeepSeek, Qwen, Ollama
Chinese providers: Baidu (文心), Alibaba (通义), Zhipu (智谱), Xunfei (讯飞)

The framework includes hardware acceleration through Flash Attention, quantization (BitsAndBytes 8-bit/4-bit, GPTQ), and CUDA support (11.8, 12.1, 12.4).

For model configuration, see Service-oriented Multi-model Management Framework (SMMF).

Sources: README.md180-292 README.md-zh.md191-318 Diagram 3 from provided context

Technology Stack and Dependencies

Core Technologies

Key Dependencies by Package

dbgpt-core: sqlalchemy, pydantic, fastapi
dbgpt-ext: Optional extras for storage (storage_milvus, storage_chromadb, graph_rag)
dbgpt-accelerator: vllm, flash-attn, bitsandbytes, gptq
dbgpt-app: fastapi, uvicorn, sqlalchemy

The default installation uses SQLite for metadata storage (no external database required). Production deployments can switch to MySQL or PostgreSQL via configuration.

Sources: packages/dbgpt-core/src/dbgpt/storage/metadata/db_storage.py1-49 docs/docs/quickstart.md27-28 Diagram 5 from provided context

Entry Points and Configuration

Main Application Entry

The primary entry point is dbgpt_server.py:

packages/dbgpt-app/src/dbgpt_app/dbgpt_server.py

Start the server using:

Or directly:

Sources: docs/docs/quickstart.md137-147 docs/docs/installation/sourcecode.md106-116

Configuration Files

Configuration follows a TOML-based system in the configs/ directory:

Configuration File	Purpose	Key Sections
`configs/dbgpt-proxy-openai.toml`	OpenAI proxy setup	`[models.llms]`, `[models.embeddings]` with `api_key`
`configs/dbgpt-proxy-deepseek.toml`	DeepSeek proxy setup	LLM and embedding configuration
`configs/dbgpt-local-glm.toml`	Local GLM-4 model	HuggingFace model paths with `provider = "hf"`

Example configuration structure:

For storage configuration:

Sources: docs/docs/quickstart.md124-135 docs/docs/installation/integrations/milvus_rag_install.md27-37 docs/docs/installation/integrations/oceanbase_rag_install.md27-37

Storage Architecture

DB-GPT implements a flexible storage abstraction through StorageInterface in dbgpt-core:

The SQLAlchemyStorage class provides the core metadata persistence:

This abstraction allows seamless switching between storage backends through configuration without code changes.

For detailed storage implementations, see Storage Architecture and Databases.

Sources: packages/dbgpt-core/src/dbgpt/storage/metadata/db_storage.py21-49 Diagram 2 from provided context

Deployment Options

DB-GPT supports multiple deployment modes:

1. Source Code Deployment

Sources: docs/docs/installation/sourcecode.md14-21 docs/docs/installation/sourcecode.md69-89

2. Docker Deployment

Multi-stage Dockerfiles support different installation modes:

Base mode: Core dependencies only
OpenAI mode: Proxy model support
vLLM mode: High-throughput local inference
LLAMA.cpp mode: CPU/Metal inference
Full mode: All features

Build and run:

Sources: Diagram 5 from provided context, docs/sidebars.js127-140

3. Docker Compose

Orchestrates multiple services (database, webserver) with volume management:

The docker-compose.yml file defines service dependencies and networking.

Sources: docs/sidebars.js131-134

4. PyPI Installation

For production deployments, install via pip:

Individual packages can be installed separately:

Sources: Diagram 5 from provided context

For comprehensive deployment instructions, see Deployment and Configuration.

Getting Started

To begin using DB-GPT:

Install Dependencies: Follow the Source Code Installation guide or use Docker
Configure Model: Choose between proxy models (OpenAI, DeepSeek) or local models (GLM-4, etc.) in configuration files
Start Webserver: Run uv run dbgpt start webserver --config <config-file>
Access UI: Visit http://localhost:5670 in your browser

For detailed quickstart instructions, see Quickstart.

For development setup and contributing, see Development Guide.

Sources: docs/docs/quickstart.md1-217 README.md139-159

Additional Resources

GitHub Repository: https://github.com/eosphoros-ai/DB-GPT
Documentation: http://docs.dbgpt.cn/docs/overview/
Paper: https://arxiv.org/pdf/2312.17449.pdf
Community: https://github.com/eosphoros-ai/community

Sources: README.md49 README.md323-349

Introduction to DB-GPT

What is DB-GPT?

System Architecture Overview

High-Level Architecture

Monorepo Package Structure

Package Hierarchy

Installation and Dependency Management

Core Capabilities

RAG (Retrieval-Augmented Generation)

GBI (Generative Business Intelligence)

Multi-Agent Framework

AWEL (Agentic Workflow Expression Language)

SMMF (Service-oriented Multi-model Management Framework)

Technology Stack and Dependencies

Core Technologies

Key Dependencies by Package

Entry Points and Configuration

Main Application Entry

Configuration Files

Storage Architecture

Deployment Options

1. Source Code Deployment

2. Docker Deployment

3. Docker Compose

4. PyPI Installation

Getting Started

Additional Resources

On this page

Introduction to DB-GPT

What is DB-GPT?

System Architecture Overview

High-Level Architecture

Monorepo Package Structure

Package Hierarchy

Installation and Dependency Management

Core Capabilities

RAG (Retrieval-Augmented Generation)

GBI (Generative Business Intelligence)

Multi-Agent Framework

AWEL (Agentic Workflow Expression Language)

SMMF (Service-oriented Multi-model Management Framework)

Technology Stack and Dependencies

Core Technologies

Key Dependencies by Package

Entry Points and Configuration

Main Application Entry

Configuration Files

Storage Architecture

Deployment Options

1. Source Code Deployment

2. Docker Deployment

3. Docker Compose

4. PyPI Installation

Getting Started

Additional Resources

On this page