How 123RF Scaled Visual Search to 200M+ Assets with Zilliz Cloud

About 123RF

123RF, part of the Inmagine Group, is one of the world's largest stock content platforms — serving millions of creative professionals with a library of over 200 million images, videos, and audio files. Search is the core of the 123RF experience: every query must surface the most relevant visual content from a massive, constantly growing catalog. When rising costs and unreliable performance on OpenSearch threatened that experience, 123RF turned to Zilliz Cloud — cutting infrastructure costs by over 50%, halving query latency, and eliminating the indexing failures that had plagued their previous setup.

The Challenge

Previously, 123RF relied on OpenSearch as its primary search infrastructure. The platform was originally built around full-text keyword search, but as the AI era arrived, the team began experimenting with embedding-based semantic search to deliver more relevant results. They layered the KNN plugin onto their existing OpenSearch cluster rather than rebuilding from scratch.

That decision carried a mounting cost. Three interlocking problems eventually made the status quo untenable:

• Escalating costs: Running a KNN-enabled OpenSearch cluster at a scale of 200M+ vectors pushed monthly operating expenses into five-digit territory and kept rising.

• Unreliable performance: Latency and query throughput became unpredictable under real production traffic, degrading the search experience for end users.

• Indexing instability: Because 123RF’s library grows daily, new assets must be indexed continuously. The OpenSearch cluster experienced frequent node failures during these indexing operations, requiring ongoing DevOps intervention.

OpenSearch was not purpose-built for vector similarity search. Its KNN plugin provided a workaround, but managing it at scale created operational overhead that the team could not sustainably absorb.

Why Zilliz Cloud

When Su-Meng Yong and his team set out to find an alternative, they evaluated several dedicated vector database options such as Pinecone and Weaviate. Three criteria drove the decision:

• Scale: The solution had to handle hundreds of million vectors reliably without performance degradation.

• Cost efficiency: Some alternatives were ruled out because they would cost more to operate at 123RF’s required scale.

• Maturity and community feedback: Zilliz Cloud is a fully managed service built on the open-source Milvus vector database, which has a dynamic community.

The Solution

123RF deployed Zilliz Cloud to power two complementary search workflows:

• Text-to-image search: User queries are converted into vector embeddings, which are then matched against the indexed image library using vector similarity, returning semantically relevant results.

• Reverse image search: Users upload an image; the system generates its embedding and searches for visually similar assets across the full library.

The embedding layer uses CLIP, an open-source multimodal embedding model, which the team iterated on across two model versions with support from the Zilliz solutions team. The flexibility to use any embedding model — not a prescribed vendor model — was noted as a meaningful advantage.

A daily batch pipeline converts all new contributor submissions into embeddings and ingests them into the Zilliz Cloud cluster, keeping the index up to date without manual intervention.

Three platform capabilities proved particularly valuable during deployment:

• Dynamic scaling: The cluster can be scaled up or down based on anticipated query load, a capability not available in the previous OpenSearch setup.

• Bulk import jobs: Zilliz Cloud’s import job feature allows indexing millions to tens of millions of rows within hours, resolving the chronic indexing bottleneck that had caused node failures under OpenSearch.

• Boost Ranker (custom feature): 123RF required custom business logic in its search ranking. The Zilliz engineering team developed a Boost Ranker feature specifically for this use case, which is now running in production.

Results and Benefits

>50% Cost Reduction

The most immediate impact was financial. With help from the Zilliz team, 123RF brought its monthly search infrastructure costs down to a fraction of the original spend — a reduction of over 50%.

"Search is the heart of our platform — it's how millions of users find the right content. Moving to Zilliz Cloud didn't just cut our infrastructure costs dramatically; it gave our engineering team the confidence that search will scale with our business instead of holding it back."

— Su-Meng Yong, Engineering Team Lead, 123RF

< 50ms latency achieved

After several optimization iterations with the Zilliz team, 123RF reduced average query latency from 100ms to 30-50ms — a roughly 50% improvement — while maintaining production-level throughput and daily traffic loads.

Zero-Downtime Indexing

The node-dropping issues that plagued OpenSearch during daily content ingestion disappeared entirely. Previously, the team could not index new images into the cluster fast enough without degrading search performance for live users. Using Zilliz Cloud's bulk import capability, the team now indexes millions to tens of millions of new rows within hours — with zero impact on query performance. A daily automated pipeline converts newly submitted stock content into embeddings and ingests them into the cluster, keeping the search index up to date without manual intervention.

Operational Freedom

As a fully managed service, Zilliz Cloud eliminated the cluster management burden that had consumed the DevOps team's time. The engineering team shifted from firefighting infrastructure issues to building product features.

"It really saves both my team, and also the developers a lot of time from having to deal with a lot of problems, a lot of self-managing of the cluster." — — Su-Meng Yong, Engineering Team Lead, 123RF

What’s Next

With image search fully migrated and stable, 123RF plans to bring its video and audio search workflows to Zilliz Cloud. The team is also open to exploring LangChain or LlamaIndex integrations in the future to extend the search capabilities of their platform.