LiveGraph: A Compound AI System for Evolving Knowledge Graph Construction from Streaming Data

LiveGraph is a compound AI system for continuously updating knowledge graphs from streaming data using a formal operator algebra—five atomic operators with inverses—that enables full rollback, operator logging, and component-level error attribution. It makes incremental knowledge graph construction from heterogeneous live data sources reliable and debuggable.

Abstract

Constructing knowledge graphs from streaming data requires coordinating LLMs, embedding models, graph databases, and domain adapters—yet existing systems compose these components through ad-hoc interfaces, making them brittle to component changes and opaque to debugging. We present LiveGraph, a compound AI system that orchestrates these heterogeneous components through a formal graph evolution operator algebra comprising five atomic operators (AddNode, AddEdge, MergeEntity, UpdateRelation, ExpireEdge) with inverses enabling full rollback and operator logs allowing component-level error attribution. The algebra serves as the compositional interface: LLMs produce candidate extractions, embedding models retrieve resolution candidates, and the storage backend applies operators atomically—all through the same shared operator vocabulary. The resulting graph serves as a reasoning substrate—graph traversal discovers multi-hop relational paths unavailable to flat retrieval, and temporal validity windows enable time-aware querying over the evolution history. We evaluate on three domains—meeting conversations (AMI), geopolitical events (ICEWS18), and Wikipedia editor interactions—verifying that the operator algebra enforces temporal coherence (S_C = 1.0) across all three, and we extend the throughput evaluation to 100K events (S_C = 1.0 at 91K events/sec). Against AMI ground truth, LiveGraph achieves Entity F1 of 0.96 (GPT-4.1) and 0.85 (GPT-4.1-mini), with entity-resolution contributing +6 pp on the smaller extractor; on external benchmarks Re-DocRED and Text2KGBench, the same generic adapter yields Entity F1 of 0.72–0.77 with no task-specific tuning. On downstream QA over transcript-derived questions, graph-augmented retrieval from LiveGraph's graphs achieves 33% accuracy vs. 7% for Microsoft GraphRAG Local on identical content (a 5x gap), with the advantage holding across factual, multi-participant, and temporal question types. The modular architecture enables systematic component evaluation through swappable interfaces and modular storage backends.