u.uday / AI FIELDNOTES
Netskope projects
Netskope / RAG & AGENT SYSTEMSProduction experience

Hybrid GraphRAG: Merging Neo4j with Vector Retrieval

A production RAG application combining graph relationships, vector search and multi-agent workflows.

PythonLangGraphNeo4jCopilotKitDockerGCP
SYSTEM OVERVIEW
the pieces, connected.
Question (CopilotKit application) → Query decomposition (LangGraph agents) → Graph retrieval (Neo4j / structured sources) + Vector retrieval (Unstructured documents) → Gather relevant context (Document-search agents) → Synthesize the answer (LangGraph / application)

A public sketch of the documented components and workflow. Internal interfaces are omitted.

Download full-resolution SVG
THE PROBLEM

Where the work started.

Enterprise knowledge spans structured records and unstructured documents. Answering a question requires finding relevant information and understanding the relationships that give it context.

MY CONTRIBUTION

What I built.

  • Built a production retrieval application combining structured graph data with vector search.
  • Developed agent workflows for document search, query decomposition and answer synthesis with LangGraph and Neo4j.
  • Integrated CopilotKit into the application experience.
  • Containerized Python services and automated deployments with GCP Cloud Build and Cloud Run.
ENGINEERING APPROACH

How the system comes together.

01

Connect documents and relationships

Neo4j · graph + vector search

Architected retrieval across structured and unstructured sources, combining graph relationships with semantic document search.

02

Coordinate the work

LangGraph · CopilotKit

Developed specialized agents for query decomposition, document search and answer synthesis, with CopilotKit integrated into the application experience.

03

Deliver the application

Python · Docker · GCP

Implemented containerized Python services and automated deployments through Cloud Build and Cloud Run. Used uv to manage Python environments.

THE RESULT
3×faster retrieval across internal docs

Brought enterprise knowledge retrieval into a production application with automated cloud deployments.

Selected work from my engineering role at Netskope. Impact figures are reported from this work; the companion examples below use synthetic data.

CODE & EVALUATION

Inspect the work behind the explanation.

SYNTHETIC COMPANION EXAMPLE

Precision, recall and reciprocal rank for retrieved evidence

A new, runnable example that scores saved predictions against expected outputs. It uses fictional data and illustrates evaluation mechanics; it does not reproduce the production system or substantiate the impact figures above.

python3 evaluate_outputs.py retrieval.json report.json
Download runnable examplesRead the script Sample report

Python 3.10+ · Standard library only · Includes fixtures, metric definitions and tests

Explore more projects Discuss this project