Agentic RAG with Python

Transform your Temporal Graph into an autonomous Intelligence Agent.

While standard RAG implementations rely on passive vector retrieval, Agentic RAG empowers your LLM to actively "drive" the graph. By providing the LLM with a Python execution environment and a Raphtory graph, it can write its own traversals, calculate metrics on-the-fly, and determine its own retrieval strategy.

Core Philosophy: The LLM with a Steering Wheel

In this architecture, the LLM is not just a consumer of data; it is an analyst with access to the PersistentGraph.

Implementation Guide

To implement Agentic RAG, you expose a Python tool to your LLM (using frameworks like LangChain, LlamaIndex, or Vercel AI SDK).

1. The Python Execution Tool

Your backend should provide a secure environment where the LLM can execute Raphtory code. Here is a simplified pattern:

2. Teaching the Agent to Search (FastRP)

You can significantly boost the agent's power by teaching it how to use Raphtory's native embedding algorithms like FastRP for similarity search.

In your Tool's documentation for the LLM, provide this recipe:

Technical Patterns

Dynamic Introspection

Don't just give the LLM the graph. Give it the Metadata. Before the conversation starts, run a "Pre-flight" script to extract:

• Unique node types and edge layers.
• Property keys and sample values.
• The current temporal range of the graph.

Inject this into the System Prompt so the LLM knows exactly which properties it can filter on.

Subgraph Slicing

Instruct the agent to "Zoom in" before analyzing. A common pattern is to have the agent identify a small set of interesting "Seed Nodes" and then create a subgraph of their 2-hop neighborhood for detailed reasoning.

Comparison: Agentic vs. Standard

Next Steps