PromptAgent: An Agent Leveraging LLM Prompting for Text-to-Graph and Graph-to-Text

PromptAgent: Text-to-Graph and Graph-to-Text with LLM Prompting

We’re excited to showcase PromptAgent, an AI system that utilizes LLM prompting to convert text into knowledge graphs (text2graph) and reconstruct structured graphs back into textual descriptions (graph2text).

The HumemAI Python Package

PromptAgent is built on top of the HumemAI Python package, abstracting away complex interactions with graph databases and LLMs. While Python functions make it easy to use, the system operates under the hood with graph databases, Gremlin queries, and LLM-based prompting.

Knowledge Graphs: JanusGraph and Cassandra

For representing structured knowledge, PromptAgent uses property graphs instead of RDF, leveraging JanusGraph (an open-source graph database) with Cassandra as the backend storage system. This combination offers efficient graph traversal and scalability, making it well-suited for real-world AI applications.

LLM for Graph Processing

Since LLMs are naturally optimized for sequential text rather than structured graphs, we serialize graph structures into JSON format before passing them to an LLM. This enables transformations like:

• Text-to-Graph (text2graph): Extracting structured relationships from text.
• Graph-to-Text (graph2text): Generating meaningful descriptions from graph data.

Post-processing methods, such as regex-based parsing, help reconstruct structured knowledge from LLM outputs.

Human-Like Memory Integration

Inspired by human memory models, PromptAgent distinguishes between short-term memory and long-term memory:

• Short-term memory: Stores newly processed text as a temporary graph with a current_time property.
• Working memory: Retrieves only relevant long-term memories, filtering them based on graph distance (num_hops).
• Long-term memory storage: Tracks frequently accessed memories by incrementing a num_retrieved counter.

This approach ensures that PromptAgent remains scalable, leveraging graph-based memory retrieval rather than keeping an entire dataset in context, allowing it to handle vast amounts of knowledge efficiently.

Future Improvements

• Enhanced Prompt Optimization: Refining LLM prompts to improve knowledge extraction.
• Integration with Larger LLMs: Expanding PromptAgent to leverage advanced OpenAI APIs or alternative LLMs to reduce hallucinations.
• Interactive Chat Interface: Enabling real-time interaction with PromptAgent for assisting users in memory-intensive tasks.

Stay tuned for further updates as we refine and expand PromptAgent’s capabilities!