←SELECTED WORKS•CASE 05•ai

AI Search Optimization using RAG

A RAG layer using Qdrant vector database and dense embeddings to solve CUHK building misidentification in LLMs, specifically configured for Google Vertex AI and published in Play Lab as a CUHK place finder

CUHK PLACE FINDERQDRANT VECTOR RAG
rag.internal // cuhk-place-finder
AI Search Optimization using RAG for CUHK Places
Dense vector database indexing CUHK campus buildings with Qdrant and Google Vertex AI alignment for grounded place search.● LIVE
✦ RAG Layer: Dense Embeddings ➔ Qdrant Vector DB ➔ Vertex AI GroundingONLINE
SYSTEM // QDRANT_VECTOR

Campus Vector RAG Spatial Search

Dense Qdrant vector retrieval paired with Vertex AI to eliminate campus landmark hallucinations.

AI Search Optimization using RAG for CUHK Places

Dense vector database indexing CUHK campus buildings with Qdrant and Google Vertex AI alignment for grounded place search.

01 // PROVEN IMPACTVERIFIED
0% Grounded

Eliminates CUHK landmark hallucination by grounding Google Vertex AI with an external Qdrant vector database.

BACKEND: QDRANT VECTOR SERVER● LIVE IN PLAY LAB
02 // THE CONTEXTPROBLEM ➔ FIX
THE CHALLENGE

General LLMs struggle to identify specific CUHK buildings from photos. With scarce campus building data in standard training sets and visually similar campus structures, models frequently give wrong location answers.

THE SOLUTION

Introduced a RAG layer with Qdrant vector database and dense embeddings, specifically matching Google Vertex AI's working principles so the AI can accurately retrieve and identify CUHK places.

ZERO LONG PARAGRAPHS
WORKFLOW // RAG RETRIEVAL PIPELINE
BACKENDCLIENT SIDE & FRONTENDPromptContextIDsAPIDataQdrantBackend ServerDatabaseAPI :6333Retrieved DataBuildingTagsDescriptionPhotosSpatial metadata packageClient serverOrchestration LayerPrompt + Skills + DB ➔API lookup ➔ Chat responseDirect API QueryChat interfaceFRONTENDPromptFind campus landmarkBuilding · Tags · Description · PhotosCUHK Place Verified ✓Matched CardSkillsConfigured RulesSet Skills · GroundingLinked to Qdrant & Vertex AIInstruction PipelineVertex AIGoogle CloudMatching IDsResolves Landmark IDs ➔ ReturnsGrounding Vector Match
VECTOR SEARCH · EMBEDDING RETRIEVALRAG PIPELINE ACTIVE ✓
02 // XAVIER'S ROLE & HOW HE BUILT ITRAG ARCHITECT & SYSTEMS ENGINEER

CASE STUDY ROLE: RAG ARCHITECT & SYSTEMS ENGINEER

Xavier built a dedicated RAG layer using Qdrant as the backend vector server with a dense embedding model to index CUHK campus buildings.

A tricky engineering challenge was that different LLMs behave differently: because he used Google Vertex AI, the Qdrant configuration and retrieval formatting had to match the AI's working principles so the model can read and process the results easily.

The product is effectively a CUHK place finder that makes searching campus locations much easier, and it is published in Play Lab for anyone to try.

WHAT XAVIER DIRECTLY DELIVERED & SHIPPED
Identified and solved LLM failure modes where models misidentified similar-looking CUHK buildings from photos
Deployed Qdrant as the backend vector server with a dense embedding model building the vector database
Tuned Qdrant configuration and retrieval structures to match Google Vertex AI's specific working principles
Created a reliable CUHK place finder that makes searching campus locations significantly easier
Published the live interactive demo in Play Lab so users can test the RAG place finder firsthand
HANDS-ON OWNERSHIPDIRECT ARCHITECT & BUILDER
03 // LANGUAGES & TOOLS USED
LANGUAGES & CORE RUNTIME
PythonTypeScriptSQL
FRAMEWORKS & ENGINES
FastAPINext.jsDocker Compose
TOOLS, AI MODELS & APIS
Qdrant Vector DBGoogle Vertex AIDense Embedding Model
STORAGE & INFRASTRUCTURE
Qdrant Vector ServerPlay Lab Client
CASE STUDY TECH ECOSYSTEM
PythonQdrant Vector DBDense EmbeddingsGoogle Vertex AIFastAPIDocker ComposeTypeScript
TECH ARCHITECTURE● FULLY FUNCTIONAL
04 // HOW THE SYSTEM WORKS

HOW THE CUHK PLACE FINDER WORKS

4-STEP PIPELINE
STEP 01
01/04

Building Query or Photo

User provides a CUHK building inquiry or photo that conventional LLMs struggle to place.

StatusActive
STEP 02
02/04

Dense Embedding

A dense embedding model converts the location query into a high-dimensional vector representation.

StatusActive
STEP 03
03/04

Qdrant Vector Retrieval

Qdrant searches the vector database, formatted specifically for Google Vertex AI's working principles.

StatusActive
STEP 04
04/04

Accurate Place Result

Vertex AI reads the vector results to deliver the exact CUHK place. Live in Play Lab for visitors to try.

StatusActive