Neo4j Semantic Search for LLM Applications

Overview

Built at Goodfolio (London, remote) as part of a UK-based product team: a data pipeline that collects professional profile data only from users who voluntarily share it with explicit permission, models it in a Neo4j graph, and lets LLM applications retrieve relevant insights through semantic search.

Consent-based data collection

Users shared professional profile data (for example from LinkedIn) through Google Forms and other Google Workspace tools, with explicit permission. The data flowed into Google Cloud Platform, where LLMs analyzed it to power end-to-end service delivery.

flowchart TB
    A[Users share data with explicit permission] --> B[Google Forms / Workspace]
    B --> C[GCP pipeline]
    C --> D[(Neo4j graph)]
    D --> E[LLM application: semantic search]

Graph-backed semantic search

Neo4j graph databases model the relationships in the collected data so an LLM application can retrieve relevant context for a question instead of relying on plain keyword or vector lookup alone.

Prompting

Few-shot and structured prompting raised LLM output accuracy in production.

Results