AI BriefingLexifina Knowledge Graph Cuts Agent Latency by 73% and Token Usage by 92%
·2026.10.20 09:00
Key point
A hybrid rules-and-LLM graph construction method reduces agent input tokens by 92% and model calls by 66% while maintaining 100% accuracy on firm-wide queries.
Details
- Lexifina demonstrates that providing agents with a firm-wide knowledge graph reduces answer time by 73%, input tokens by 92%, and model calls by 66% compared to file-based retrieval.
- The graph is built using a hybrid approach where general drafting rules extract most facts at zero token cost, and a small model (gpt-5.4-mini) processes only the 1.6% of sentences where rules fail, costing just $0.047 for 245 documents.
- In a benchmark of five firm-wide questions, the graph-enabled agent achieved 100% accuracy with 102k input tokens and 30 tool calls, whereas the file-based agent used 1.35M tokens and 205 tool calls for 99% accuracy.
- Existing clause metadata was found to be a poor source for graph construction, missing all liability caps and several deadlines, while costing $3.15 in the test run compared to the $0.047 cost of the targeted model pass.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Uber Achieves 52% Reduction in Per-Session AI Agent Costs Through Cost Structure Analysis and Optimization