RAG (Retrieval-Augmented Generation)

RAG enhances LLM responses by retrieving relevant context from your own data before generating an answer. DB-GPT provides a comprehensive RAG framework with multiple indexing and retrieval strategies, and runs knowledge-base chat as an agentic RAG loop.

Two phases: indexing and conversation

INDEXING  (document-sync time)        CONVERSATION  (chat time)
─────────────────────────────         ──────────────────────────
upload → chunk → indexes              question
                       │                       │
        vector / keyword /                  agentic RAG loop:
        knowledge-graph                     rewrite → retrieve
        (+ structural tree,                 (maybe repeat) →
         + code graph for code)             fuse + rerank →
                                             cited answer

These two phases are decoupled — indexing runs once at sync time; chat only retrieves and never re-indexes.

How the conversation works (agentic RAG)

DB-GPT does not do a single retrieve-then-generate. An agent drives the loop:

The agent can rewrite the query, retrieve multiple times, fuse and rerank results, and produce a cited answer — letting it handle complex or multi-part questions that one-shot RAG cannot. Full detail: Agentic RAG Conversation Principles.

Indexes

All indexes operate on the same chunks, so chunking quality is the foundation.

Index	What it gives you	Built
Vector	semantic similarity via embeddings + cosine	at sync time
Keyword (BM25)	exact term matching	at sync time
Knowledge Graph	graph traversal over entities, document structure, headings, and code	at sync time
Structural	markdown-header tree / parent-child navigation	at query time (from chunk heading metadata)
Code graph	code AST as function / class nodes (for code files / git repos)	at sync time

tip

The Knowledge Graph index is a family of graphs: LLM-extracted triplets, a document–paragraph structure graph, a Markdown heading-hierarchy graph, and a code AST graph (parsed with tree-sitter). See Knowledge Base Indexing Principles.

Supported file formats

Upload and process a wide variety of document formats:

• Documents: PDF, Word (.docx), Markdown, TXT
• Spreadsheets: Excel (.xlsx), CSV
• Web: HTML, URLs
• Code: Python, Java, JavaScript, TypeScript, Go, Rust, C/C++ — parsed into the code graph

Quick start with RAG

1. Open the DB-GPT Web UI
2. Navigate to Knowledge Base in the sidebar
3. Create a new knowledge base (choose index methods: Vector / Knowledge Graph / Full Text — combinable)
4. Upload your documents
5. Wait for processing to complete
6. Start chatting with your knowledge base (runs the agentic RAG loop)

For programmatic access, see the RAG Cookbook.

What's next

• Knowledge Base Indexing Principles — how a document becomes searchable
• Agentic RAG Conversation Principles — how a question becomes a cited answer
• Knowledge Base UI — Manage knowledge bases in the Web UI
• Graph RAG — Knowledge graph-based retrieval
• RAG Module — Deep dive into the RAG framework
• RAG Development Guide — Build RAG apps programmatically