DIV-1 · Divanshu Sharma

Work / Fig. 06 · Applied LLM product · 2026

Graph Data Explorer AI

765 nodes / 877 edges — typed knowledge graph from 19 source tables.

Fig. 06

Graph Data Explorer AI

Applied LLM product · 2026

Problem

21,393 SAP order-to-cash records across 19 flat tables — relationships a business user can neither see nor query without SQL and schema knowledge.

Approach

Transformed the tables into a typed knowledge graph (765 nodes, 877 edges) served by FastAPI with SSE streaming and a React Flow visual layer; built a two-stage LLM pipeline (natural language → sandboxed graph traversal → natural-language explanation) wrapped in layered guardrails: keyword pre-filters, sandboxed execution, forbidden-token checks.

Results

  • 765 nodes / 877 edges

    typed knowledge graph from 19 source tables

  • Two-stage LLM query engine

    NL → traversal → NL, fully sandboxed

  • 3 guardrail layers

    pre-filters, sandbox, forbidden-token checks