What challenges did 123RF face before AI implementation?

Translating English-only content titles and metadata into 15 languages at massive scale while balancing translation quality and cost. Google Translate was too expensive and some LLM options were costly or inconsistent. The content contained idioms, named entities, and domain-specific jargon that simple translation tools handled poorly.

What was the impact of the AI solution?

Achieved a 95% reduction in translation costs. Rolled out translations to 9 languages quickly with plans to cover all 15. Improved translation quality and handling of idioms, named entities, and technical jargon. Reduced time-to-publish multilingual content.

Source

Scaled productionEvidence: Medium50/100

123RF reduces translation costs 95% using Amazon Bedrock and dynamic prompt sampling

Use case typeMultilingual communicationUpdated Jun 13, 2026

123RF built an LLM translation assistant to translate English-only content titles and metadata into multiple languages at scale. The solution uses Amazon Bedrock with Claude 3 Haiku, embeddings, and a vector store for retrieval augmented generation with dynamic prompt sampling and K-shot examples. The team combined prompt engineering, hybrid similarity search, and reusable translation pairs to improve quality and reduce cost.

Organization: 123RF
Industry: Tech & Comms
Location: Malaysia
Published: November 2024

Reported outcomes

−95%

costCost savings

Strategic outcomes

Cost efficiencyReduced translation costsMarket & geographic expansionRolled out multilingual translationsNew product / capabilityImproved translation qualitySpeed & agilityReduced time to publish content

Catalog median for cost savings deployments: −40% across 177 reported metrics. Compare benchmarks →

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

Cost: 95% decrease

AWS Machine Learning BlogNov 25, 2024Blog postInferred claimMedium evidence strength

Achieved a 95% reduction in translation costs.

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: 123RF
Provider: AWS
Maturity: Scaled Production
Linked source: AWS Machine Learning Blog

123RF built an LLM translation assistant to translate English-only content titles and metadata into multiple languages at scale

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 3 of 3

1Translation Automation
2Multilingual Content Discovery
3Retrieval Augmented Generation

Translating English-only content titles and metadata into 15 languages at massive scale while balancing translation quality and cost.
Google Translate was too expensive and some LLM options were costly or inconsistent.
The content contained idioms, named entities, and domain-specific jargon that simple translation tools handled poorly.

Built an AI language translation assistant using role prompting, separation of data and templates, scratchpad reasoning, and K-shot examples.
Implemented dynamic prompting with a vector database of prior high-quality translation pairs and hybrid similarity search to retrieve relevant examples.
Stored source and translated chunks with embeddings and metadata for each target language, then injected matched examples into prompts for Amazon Bedrock translation.

Technologies

Amazon Bedrock Claude 3 Haiku Amazon Titan Amazon OpenSearch Service Amazon Bedrock Prompt Management

Achieved a 95% reduction in translation costs.
Rolled out translations to 9 languages quickly with plans to cover all 15.
Improved translation quality and handling of idioms, named entities, and technical jargon.
Reduced time-to-publish multilingual content.

Architecture

123RF built an LLM translation pipeline on Amazon Bedrock using Claude 3 Haiku and embedding models to create a language-specific vector database of prior translations. For each new translation task, the system used hybrid similarity search to retrieve relevant source/translation examples and dynamically inject them into prompts, combining role prompting, separation of instructions and data, scratchpad reasoning, and K-shot examples to improve quality and lower cost.

Sources & evidence1

Evidence: Medium50/100Evidence strength