TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

Knowledge Graph Verification

1 post · Nov 2023 – Jul 2026

What this trail covers

Grab has evolved its knowledge graph verification process by operationalizing the search interface as a validation environment. This feedback-driven framework injects unverified candidate edges generated by LLMs into live search results and recommendation carousels. User micro-interactions are tracked and aggregated to calculate a confidence score, which is then used to promote verified edges or prune hallucinated ones. This system ensures the accuracy and integrity of the knowledge graph in dynamic catalogs.

All Grab trails
  1. Crowdsourced taxonomy verification: A feedback-driven framework for refining knowledge graph relationships via online search interactionsThis post details the evolution of Grab's knowledge graph verification system by introducing a feedback-driven engine that operationalizes the search interface. It describes the architecture comprising a KG core with verified and candidate edges, a search and injection module for exploration vs. exploitation, a behavior tracking module for granular user interaction logging, and a verification and refinement engine for updating the KG topology. The case study on food delivery taxonomy highlights the promotion and pruning of candidate edges based on user signals.
  2. Graph modelling guidelinesThis post provides a step-by-step guide to graph modelling, covering domain definition, entity and relationship identification, property establishment, model selection (property graph vs. RDF), schema development, data import, implementation using graph databases, and analysis/visualization. It specifically uses a social network graph as an example and highlights the suitability of the property graph model for such applications.