Production-Grade Multimodal Search Infrastructure

Text + Vision AI
Search Platform

Upload a product photo and find it instantly. YOLOv11-World detects objects, SigLIP 2 encodes vision embeddings, BGE-M3 handles text — fused via RRF for unmatched hybrid search accuracy.

YOLOv11-World| SigLIP 2| BGE-M3| Qdrant| FastAPI| PostgreSQL
Key Differentiator

Multimodal Vision Search

Upload any product photo. Our AI detects objects with YOLO-World, encodes them with SigLIP 2, and fuses results with text search via Reciprocal Rank Fusion.

YOLO-World Detection

Open-vocabulary object localization identifies products in any photo — no pre-trained classes needed. Outputs bounding boxes for each detected item.

IMG
bottle 0.94
box 0.87
can 0.82

SigLIP 2 Embeddings

State-of-the-art vision encoder maps product images into a 1152-dim embedding space. Cross-modal alignment ensures text and image queries share the same vector space.

dim-0
dim-1
dim-2
dim-3
...

1152 dimensions · so400m-patch14-384

Smart Crop Modes

crop=true: YOLO detects → crops each object → embeds individually for precise per-item matching. crop=false: Embeds the full image for scene-level similarity.

crop=true
OBJ
OBJ

Per-object embed

crop=false
FULL IMG

Scene-level embed

Reciprocal Rank Fusion

Text results (BGE-M3 dense + sparse) and image results (SigLIP 2 vision) are independently ranked, then merged via RRF to produce a single unified result set — capturing both semantic meaning and visual similarity.

Text
BGE-M3
+
Vision
SigLIP 2
Fused
RRF

V1 vs V2: See the Difference

V1 delivers fast keyword search. V2 adds a full NLP pipeline — NER, transliteration, spellcheck, synonym expansion, and reranking.

V1 — Plain Text Search
xaxox
mode: text 42ms · 0 results

No results found for "xaxox"

Latin input not understood

V2 — Advanced Extended Search SOTA
xaxox
translit: խաղող synonym: խաղող NER: none
mode: text 87ms · 24 results
IMG

Խաղող պրեմիում 1 կգ

Մրգերscore: 0.94
IMG

Խաղող Պերու 1 կգ

Մրգերscore: 0.89
IMG

Խաղող սեվ անկոշտ 1 կգ

Մրգերscore: 0.85

Three Pillars of Intelligence

Each layer works in concert to deliver results that feel like reading the shopper's mind.

Multimodal Fusion

SigLIP 2 vision embeddings + BGE-M3 text embeddings + YOLO object detection. Snap a photo and find it instantly via server-side RRF fusion.

SigLIP 2 BGE-M3 YOLO v11 RRF

Native Script Transliteration

Algorithmic "Arminglish" transliteration handles Latin-typed Armenian (xaxox → խաղող), slang dictionaries, recursive synonym expansion, and agglutinative stemming.

HY / RU / EN Arminglish Stemming NER

Strategic Searchandising

Dynamic brand and category boosts, trending scores, merchandising controls, real-time analytics, and "Did you mean?" spell-check to maximize conversion rates.

Boosts Analytics Spellcheck Reranking

Built for Speed

Every millisecond counts. Engineered for sub-50ms responses at scale.

< 15ms

Autocomplete Latency

< 45ms

Text Search (V1)

300K+

Product Embeddings

99.9%

Uptime SLA

10-Stage Search Pipeline

1Normalize query (Unicode, lowercasing)
2Spellcheck — "Did you mean?" correction
3NER extraction (brand / category filters)
4Recursive expansion (slang → translit → synonyms)
5BGE-M3 dense + sparse embedding
6Qdrant filter construction (NER + facets)
7Hybrid RRF search (dense + sparse prefetch)
8Reranking + searchandising boosts
9Async analytics logging (fire-and-forget)
10Response with full NLP transparency
Business Intelligence

Advanced Searchandising & Business Logic

Take absolute control over your catalog’s visibility. Blend state-of-the-art AI semantics with your strategic business goals.

Entity-Level Control

Boost at the Product, Brand, or Category level. Elevate high-margin collections or demote out-of-season brands instantly.

Product
×1.8
Brand
×1.5
Category
×1.2

Dynamic Trending Score

Inject a real-time trending_score (0.0 to 1.0) to dynamically push viral or fast-selling products to the top without breaking search relevance.

24h ago score: 0.87 now

Multiplicative Scoring Engine

Our proprietary RRF seamlessly multiplies pure vector similarity with your custom business weights, ensuring relevant but profitable results.

similarity
0.92
×
boost
1.50
×
trend
1.87
=
final
2.58

Zero-Latency Injection

All business rules and trending scores are cached in-memory, guaranteeing sub-millisecond ranking adjustments without hitting the database.

DB Query
12.4 ms
In-Memory
0.03 ms

413× faster · Zero database round-trips

The Scoring Formula

Every result is scored through a multiplicative pipeline that preserves AI relevance while amplifying your business priorities.

final = sim × (1 + boost) × (1 + trend)

Ready to explore?

Access the full management dashboard to test search, manage synonyms, and monitor analytics.

CityIT Logo

Enterprise Integrations & Custom Deployments

Ready to supercharge your e-commerce search? For enterprise integrations and custom deployments, contact our core engineering team.

Contact CityIT Engineering