GraphSight Pro
Multi-agent memory tracer & knowledge graph visualizer
A production-grade agent tracing framework with smart lexical overlap detection, LangGraph & LlamaIndex adapters, and interactive graph diffs.
Problem
AI agent frameworks frequently suffer from retrieval blind spots, tool execution opacity, and redundant context loading that cannot be tracked or audited in production.
Approach
Engineered an open-source observability engine that models multi-agent execution steps as interactive directed acyclic graphs (DAGs), computing real-time lexical overlap to detect whether retrieved entities were actually utilized in responses.
What shipped
- Published official Python SDK packages (graphsight-pro and graphsight-pro-langgraph) to PyPI.
- Designed sublinear TF + bigram lexical similarity algorithms for accurate hallucination & retrieved-vs-used detection.
- Deployed full-stack interactive canvas with real-time score threshold filtering and one-click JSON trace export.
Stack
Python
React
TypeScript
FastAPI
Docker