S&P Global's AI unit Kensho built 'Grounding', a LangGraph-based multi-agent framework that routes natural language queries to specialized financial data agents with citation-backed responses.
Kensho, S&P Global's AI innovation arm, developed a production multi-agent system called 'Grounding' using LangGraph. The framework uses a central router to direct natural language queries to specialized Data Retrieval Agents (DRAs) across divisions including equity research, fixed income, and macroeconomics. It returns citation-backed responses grounded in verified S&P Global datasets. The architecture separates routing logic from retrieval logic and includes custom evaluation protocols measuring exact-match and tool-calling accuracy.
Kensho's architecture solves a real production problem: how to route natural language queries across heterogeneous structured financial datasets without baking retrieval logic into every agent. The separation between a central router and domain-specific DRAs is a clean pattern that scales as data sources multiply. The custom evaluation protocol — exact-match agent selection plus tool-calling accuracy — is the part most teams skip and then regret.
Use LangGraph to prototype a router + two specialized sub-agents on your own domain this week: measure tool-calling accuracy before you add a third agent. If accuracy drops below 80%, your routing prompt needs domain-specific few-shot examples.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Install with: pip install langgraph langchain-openai and open a new Python file
Paste this skeleton: from langgraph.graph import StateGraph; define two nodes — 'router' that classifies query intent and 'retriever' that returns a mock structured response with a citation field; wire them with graph.add_edge('router', 'retriever')
Run the graph with a sample query like 'What is Apple's P/E ratio?' and inspect whether the router node correctly classifies it before passing to retrieval
Tags