Knowledge Graphs

2 min read

Archestra can automatically ingest documents uploaded via Chat into a knowledge graph. This enables graph-based retrieval augmented generation (GraphRAG) across all your organization's documents.

How It Works

When users upload documents through the Chat interface, Archestra automatically:

Extracts text content from supported file types
Sends the content to the configured knowledge graph provider
The provider indexes the document for later retrieval

This happens asynchronously in the background without blocking chat responses.

Supported File Types

Text-based documents that can be meaningfully indexed:

Text files: .txt, .md, .markdown
Data formats: .json, .csv, .xml, .yaml, .yml
Web files: .html, .htm
Code files: .js, .ts, .jsx, .tsx, .py, .java, .c, .cpp, .h, .hpp, .rs, .go, .rb, .php, .sh, .bash, .sql, .graphql, .css, .scss, .less

Binary files (images, PDFs, etc.) are not currently supported.

Configuration

Enable the feature by setting environment variables. See Deployment - Knowledge Graph Configuration for details.

LightRAG Provider

LightRAG combines vector similarity search with graph-based retrieval for more accurate and contextual results.

ARCHESTRA_KNOWLEDGE_GRAPH_PROVIDER=lightrag
ARCHESTRA_KNOWLEDGE_GRAPH_LIGHTRAG_API_URL=http://lightrag:9621
ARCHESTRA_KNOWLEDGE_GRAPH_LIGHTRAG_API_KEY=your-api-key  # Optional

LightRAG requires:

A running LightRAG API server
Neo4j for graph storage
A vector database (e.g., Qdrant) for embeddings

Using the Knowledge Graph

Once configured, documents are automatically ingested. There are two ways to query the knowledge graph from agents:

Built-in Query Tool (Recommended)

Archestra includes a built-in query_knowledge_graph tool. To use it:

Go to MCP Catalog and find "Archestra"
Assign the query_knowledge_graph tool to your profile
The tool will be available to agents using that profile

The tool is also automatically assigned to new profiles when a knowledge graph provider is configured.

External MCP Server

Alternatively, add the LightRAG MCP server to your profiles for direct LightRAG access.

Query Modes

The query_knowledge_graph tool supports different query modes:

Mode	Description	Best For
`hybrid`	Combines local and global context (default)	General queries
`local`	Uses only local context from the knowledge graph	Specific document lookups
`global`	Uses global context across all documents	Broad topic exploration
`naive`	Simple RAG without graph-based retrieval	Basic similarity search