Knowledge Base

Build and manage knowledge bases for Retrieval-Augmented Generation (RAG). Upload documents, configure retrieval, and use them in chat.

Creating a knowledge base

Step 1 — Navigate to Knowledge

Click Knowledge in the sidebar to open the knowledge management page.

Step 2 — Create a new knowledge base

1. Click Create (or the + button)
2. Fill in:
   • Name — A descriptive name for the knowledge base
   • Description — Brief description of the content
   • Embedding Model — The embedding model to use for vectorization (must match your configured embedding)
3. Click Create

Step 3 — Upload documents

1. Open the knowledge base you just created
2. Click Upload to add documents
3. Select one or more files
4. Wait for processing to complete (chunking, embedding, and indexing)

:::info Supported file formats

Using a knowledge base in chat

1. Go to Chat and create a new conversation
2. Select Chat Knowledge mode
3. Choose your knowledge base from the dropdown
4. Ask questions — the LLM will use your documents as context

Knowledge base settings

Each knowledge base has configurable settings:

:::tip Tuning retrieval

• Large documents: Increase chunk size to preserve context
• Precise answers: Increase Top K and lower the score threshold
• Noisy results: Raise the score threshold :::

Storage types

DB-GPT supports multiple vector storage backends:

To use a non-default backend, add the corresponding extra to your install command:

uv sync --all-packages --extra "storage_milvus" ...

Advanced features

DB-GPT supports knowledge graphs for structured retrieval:

• Extracts entities and relationships from documents
• Enables graph-based queries alongside vector search
• Useful for complex domain knowledge with interconnected concepts

See Graph RAG for setup instructions.

For hybrid retrieval combining vector and keyword search:

uv sync --all-packages --extra "rag_bm25" ...

This enables BM25 indexing alongside vector embeddings for improved recall.

Managing knowledge bases

:::warning Deleting is permanent Deleting a knowledge base removes all associated vector embeddings and indexed data. The original uploaded files are not recoverable. :::

Next steps