How Ring scales international buyer help with Amazon Bedrock Information Bases

This publish is cowritten with David Kim, and Premjit Singh from Ring.

Scaling self-service help globally presents challenges past translation. On this publish, we present you the way Ring, Amazon’s house safety subsidiary, constructed a production-ready, multi-locale Retrieval-Augmented Era (RAG)-based help chatbot utilizing Amazon Bedrock Information Bases. By eliminating per-Area infrastructure deployments, Ring lowered the price of scaling to every extra locale by 21%. On the similar time, Ring maintained constant buyer experiences throughout 10 worldwide Areas.

On this publish, you’ll learn the way Ring applied metadata-driven filtering for Area-specific content material, separated content material administration into ingestion, analysis and promotion workflows, and achieved value financial savings whereas scaling up. The structure described on this publish makes use of Amazon Bedrock Information Bases, Amazon Bedrock, AWS Lambda, AWS Step Features, and Amazon Easy Storage Service (Amazon S3). Whether or not you’re increasing help operations internationally or seeking to optimize your current RAG structure, this implementation supplies sensible patterns you possibly can apply to your personal multi-locale help programs.

The help evolution journey for Ring

Buyer help at Ring initially relied on a rule-based chatbot constructed with Amazon Lex. Whereas useful, the system had limitations with predefined dialog patterns that couldn’t deal with the various vary of buyer inquiries. Throughout peak durations, 16% of interactions escalated to human brokers, and help engineers spent 10% of their time sustaining the rule-based system. As Ring expanded throughout worldwide locales, this method grew to become unsustainable.

Necessities for a RAG-based help system

Ring confronted a problem: easy methods to present correct, contextually related help throughout a number of worldwide locales with out creating separate infrastructure for every Area. The group recognized 4 necessities that might inform their architectural method.

1. International content material localization

The worldwide presence of Ring required greater than translation. Every territory wanted Area-specific product info, from voltage specs to regulatory compliance particulars, supplied via a unified system. Throughout the UK, Germany, and eight different locales, Ring wanted to deal with distinct product configurations and help situations for every Area.

2. Serverless, managed structure

Ring needed their engineering group targeted on bettering buyer expertise, not managing infrastructure. The group wanted a completely managed, serverless resolution.

3. Scalable information administration

With a whole lot of product guides, troubleshooting paperwork, and help articles continuously being up to date, Ring wanted vector search expertise that might retrieve exact info from a unified repository. The system needed to help automated content material ingestion pipelines in order that the Ring content material group may publish updates that might turn out to be obtainable throughout a number of locales with out guide intervention.

4. Efficiency and value optimization

The typical end-to-end latency requirement for Ring was 7–8 seconds and efficiency evaluation revealed that cross-Area latency accounted for lower than 10% of complete response time. This discovering allowed Ring to undertake a centralized structure slightly than deploying separate infrastructure in every Area, which lowered operational complexity and prices.

To handle these necessities, Ring applied metadata-driven filtering with content material locale tags. This method serves Area-specific content material from a single centralized system. For his or her serverless necessities, Ring selected Amazon Bedrock Information Bases and Lambda, which eliminated the necessity for infrastructure administration whereas offering computerized scaling.

Overview of resolution

Ring designed their RAG-based chatbot structure to separate content material administration into two core processes: Ingestion & Analysis and Promotion. This two-phase method permits Ring to keep up steady content material enchancment whereas retaining manufacturing programs secure.

Ingestion and analysis workflow

Determine 1: Structure diagram exhibiting the Ring ingestion and analysis workflow with Step Features orchestrating each day information base creation, analysis, and high quality validation utilizing Information Bases and S3 storage.

1. Content material add – The Ring content material group uploads help documentation, troubleshooting guides, and product info to Amazon S3. The group structured the S3 objects with content material in encoded format and metadata attributes. For instance, a file for the content material “Steps to Change the doorbell battery” has the next construction:

{
	"properties": {
		"slug": "abcde",
		"contentLocale": "en-GB",  # distinctive identifier
		"sourceFormat": "md",      # locale info
		"metadataAttributes": {
			"group": "Service",
			"slug": "abcde",
			"contentLocale": "en-GB"
		},
		"content material": "U3RlcHMgdG8gUmVwbGFjZSB0aGUgZG9vcmJlbGwgYmF0dGVyeTo= 
                VXNlIHRoZSBpbmNsdWRlZCBzZWN1cml0eSBzY3Jld2RyaXZlciB0byByZW1vdmUgd
                GhlIHNlY3VyaXR5IHNjcmV3IGxvY2F0ZWQgb24gdGhlIGJvdHRvbSBvZiB0aGUgZm
                FjZXBsYXRlCgpSZW1vdmUgdGhlIGZhY2VwbGF0ZSBieSBwcmVzc2luZyBpbiBvbiB
                0aGUgc2lkZXMgYW5kIGNhcmVmdWxseSBwdWxsaW5nIGl0IG91dCBhbmQgb2ZmCgpS
                ZW1vdmUgdGhlIGJhdHRlcnkgZnJvbSB0aGUgZG9vcmJlbGwKCkNvbm5lY3QgdGhlI
                GNoYXJnaW5nIGNhYmxlIHRvIHRoZSBiYXR0ZXJ5J3MgY2hhcmdpbmcgcG9ydAoKQ2h
                hcmdlIHVudGlsIG9ubHkgdGhlIGdyZWVuIGxpZ2h0IHJlbWFpbnMgbGl0ICh3aGlsZ
                SBjaGFyZ2luZywgeW91J2xsIHNlZSBib3RoIGEgc29saWQgZ3JlZW4gYW5kIGFtYmV
                yIGxpZ2h0KQoKUmUtaW5zZXJ0IHRoZSBjaGFyZ2VkIGJhdHRlcnkgaW50byB0aGUgZ
                G9vcmJlbGwKCkRlLWF0dGFjaCB0aGUgZmFjZXBsYXRlCgpTZWN1cmUgd2l0aCB0aGU
                gc2VjdXJpdHkgc2NyZXc=    # base64 encoded 
	}
}

2. Content material processing – Ring configured Amazon S3 bucket occasion notifications with Lambda because the goal to robotically course of uploaded content material.

3. Uncooked and processed content material storage
   
   The Lambda perform performs two key operations:
   • Copies the uncooked knowledge to the Information Base Archive Bucket
   • Extracts metadata and content material from uncooked knowledge, storing them as separate recordsdata within the Information Base Supply Bucket with contentLocale classification (for instance, {locale}/Service.Ring.{Upsert/Delete}.{unique_identifier}.json)

   For the doorbell battery instance, the Ring metadata and content material recordsdata have the next construction:

   {locale}/Service.Ring.{Upsert/Delete}.{unique_identifier}.metadata.json

{
	"metadataAttributes" : {
		"group": "Service",
		"slug": "abcde",
		"contentLocale": "en-GB"
	}
}

{locale}/Service.Ring.{Upsert/Delete}.{unique_identifier}.json

{
	"content material": "Steps to Change the doorbell battery:
	Use the included safety screwdriver to take away the safety screw positioned on the underside of the faceplate
	Take away the faceplate by urgent in on the perimeters and punctiliously pulling it out and off
	Take away the battery from the doorbell
	Join the charging cable to the battery's charging port
	Cost till solely the inexperienced mild stays lit (whereas charging, you may see each a strong inexperienced and amber mild)
	Re-insert the charged battery into the doorbell
	Re-attach the faceplate
	Safe with the safety screw
}

4. Every day Knowledge Copy and Information Base Creation

Ring makes use of AWS Step Features to orchestrate their each day workflow that:

• Copies content material and metadata from the Information Base Supply Bucket to Knowledge Supply (Model)
• Creates a brand new Information Base (Model) by indexing the each day bucket as knowledge supply for vector embedding

Every model maintains a separate Information Base, giving Ring impartial analysis capabilities and simple rollback choices.

5. Every day Analysis Course of

The AWS Step Features workflow continues utilizing analysis datasets to:

• Run queries throughout Information Base variations
• Check retrieval accuracy and response high quality to check efficiency between variations
• Publish efficiency metrics to Tableau dashboards with outcomes organized by contentLocale

6. High quality Validation and Golden Dataset Creation

Ring makes use of the Anthropic Claude Sonnet 4 massive language mannequin (LLM)-as-a-judge to:

• Consider metrics throughout Information Base variations to establish the best-performing model
• Examine retrieval accuracy, response high quality, and efficiency metrics organized by contentLocale
• Promote the highest-performing model to Knowledge Supply (Golden) for manufacturing use

This structure helps rollbacks to earlier variations for as much as 30 days. As a result of content material is up to date roughly 200 instances per week, Ring determined to not preserve variations past 30 days.

Promotion workflow: customer-facing

Determine 2: Structure diagram exhibiting the Ring manufacturing chatbot system the place buyer queries circulation via AWS Lambda to retrieve context from Information Bases and generate responses utilizing basis fashions

1. Buyer interplay – Prospects provoke help queries via the chatbot interface. For instance, a buyer question for the battery substitute situation appears to be like like this:

{
	"textual content": "How can I exchange the doorbell battery?",
	"market": "en-GB"
}

2. Question orchestration and information retrieval

Ring configured Lambda to course of buyer queries and retrieve related content material from Amazon Bedrock Information Bases. The perform:

• Transforms incoming queries for the RAG system
• Applies metadata filtering with contentLocale tags utilizing equals operator for exact Regional content material concentrating on
• Queries the validated Golden Knowledge Supply to retrieve contextually related content material

Right here’s the pattern code Ring makes use of in AWS Lambda:

## Metadata Filtering for Regional Content material Focusing on

num_results = 10
market = "en-GB"
knowledge_base_id = "A2BCDEFGHI"
user_text = "How can I exchange the doorbell battery?"

# Configure Regional content material filtering
vector_search_config = {"numberOfResults": num_results}
vector_search_config["filter"] = {
	"equals": {
		"key": "contentLocale",
		"worth": market
	}
}

# Run Amazon Bedrock Information Base search
response = boto3.shopper("bedrock-agent-runtime").retrieve(
	knowledgeBaseId=knowledge_base_id,
	retrievalQuery={"textual content": user_text},
	retrievalConfiguration={
		"vectorSearchConfiguration": vector_search_config,
	},
)

3. Response era

Within the Lambda perform, the system:

• Types the retrieved content material based mostly on relevance rating and selects the highest-scoring context
• Combines the top-ranked context with the unique buyer question to create an augmented immediate
• Sends the augmented immediate to LLM on Amazon Bedrock
• Configures locale-specific prompts for every contentLocale
• Generates contextually related responses returned via the chatbot interface

Different issues in your implementation

When constructing your personal RAG-based system at scale, contemplate these architectural approaches and operational necessities past the core implementation:

Vector retailer choice

The Ring implementation makes use of Amazon OpenSearch Serverless because the vector retailer for his or her information bases. Nevertheless, Amazon Bedrock Information Bases additionally helps Amazon S3 Vectors as a vector retailer possibility. When selecting between these choices, contemplate:

• Amazon OpenSearch Serverless: Gives superior search capabilities, real-time indexing, and versatile querying choices. Greatest suited to functions requiring advanced search patterns or whenever you want extra OpenSearch options past vector search.
• Amazon S3 vectors: Presents a cheaper possibility for easy vector search use instances. S3 vector shops present computerized scaling, built-in sturdiness, and could be extra economical for large-scale deployments with predictable entry patterns.

Along with these two choices, AWS helps integrations with different knowledge retailer choices, together with Amazon Kendra, Amazon Neptune Analytics, and Amazon Aurora PostgreSQL. Consider your particular necessities round question complexity, value optimization, and operational wants when deciding on your vector retailer. The prescriptive steering supplies a very good place to begin to judge vector shops in your RAG use case.

Versioning structure issues

Whereas Ring applied separate Information Bases for every model, you would possibly contemplate an alternate method involving separate knowledge sources for every model inside a single information base. This technique leverages the x-amz-bedrock-kb-data-source-id filter parameter to focus on particular knowledge sources throughout retrieval:

vector_search_config["filter"] = {
	"equals": {
		"key": "x-amz-bedrock-kb-data-source-id",
		"worth": ''
		}
	}
# Execute Bedrock Information Base search
response = boto3.shopper("bedrock-agent-runtime").retrieve(
	knowledgeBaseId=knowledge_base_id,
	retrievalQuery={"textual content": user_text},
	retrievalConfiguration={
		"vectorSearchConfiguration": vector_search_config,
	},
)

When selecting between these approaches, weigh these particular trade-offs:

• Separate information bases per model (the method that Ring makes use of): Gives knowledge supply administration and cleaner rollback capabilities, however requires managing extra information base situations.
• Single information base with a number of knowledge sources: Reduces the variety of information base situations to keep up, however introduces complexity in knowledge supply routing logic and filtering mechanisms, plus requires sustaining separate knowledge shops for every knowledge supply ID.

Catastrophe restoration: Multi-Area deployment

Think about your catastrophe restoration necessities when designing your RAG structure. Amazon Bedrock Information Bases are Regional sources. To realize sturdy catastrophe restoration, deploy your full structure throughout a number of Areas:

• Information bases: Create Information Base situations in a number of Areas
• Amazon S3 buckets: Preserve cross-Area copies of your Golden Knowledge Supply
• Lambda capabilities and Step Features workflows: Deploy your orchestration logic in every Area
• Knowledge synchronization: Implement processes to maintain content material synchronized throughout Areas

The centralized structure serves its visitors from a single Area, prioritizing value optimization over multi-region deployment. Consider your personal Restoration Time Goal (RTO) and Restoration Level Goal (RPO) necessities to find out whether or not a multi-Area deployment is important in your use case.

Basis mannequin throughput: Cross-Area inference

Amazon Bedrock basis fashions are Regional sources with Regional quotas. To deal with visitors bursts and scale past single-Area quotas, Amazon Bedrock helps cross-Area inference (CRIS). CRIS robotically routes inference requests throughout a number of AWS Areas to extend throughput:

CRIS: Routes requests solely inside particular geographic boundaries (equivalent to inside the US or inside the EU) to fulfill knowledge residency necessities. This could present as much as double the default in-Area quotas.

International CRIS: Routes requests throughout a number of business Areas worldwide, optimizing obtainable sources and offering greater mannequin throughput past geographic CRIS capabilities. International CRIS robotically selects the optimum Area to course of every request.

CRIS operates independently out of your Information Base deployment technique. Even with a single-Area Information Base deployment, you possibly can configure CRIS to scale your basis mannequin throughput throughout visitors bursts. Notice that CRIS applies solely to the inference layer—your Information Bases, S3 buckets, and orchestration logic stay Regional sources that require separate multi-Area deployment for catastrophe restoration.

Embedding mannequin choice and chunking technique

Choosing the suitable embedding mannequin and chunking technique is essential for RAG system efficiency as a result of it straight impacts retrieval accuracy and response high quality. Ring makes use of the Amazon Titan Embeddings mannequin with the default chunking technique, which proved efficient for his or her help documentation.

Amazon Bedrock provides flexibility with a number of choices:

Embedding fashions:

• Amazon Titan embeddings: Optimized for text-based content material
• Amazon Nova multimodal embeddings: Helps “Textual content”, “Picture”, “Audio”, and “Video” modalities

Chunking methods:

When ingesting knowledge, Amazon Bedrock splits paperwork into manageable chunks for environment friendly retrieval utilizing 4 methods:

• Normal chunking: Fastened-size chunks for uniform paperwork
• Hierarchical chunking: For structured paperwork with clear part hierarchies
• Semantic chunking: Splits content material based mostly on matter boundaries
• Multimodal content material chunking: For paperwork with combined content material varieties (textual content, photographs, tables)

Consider your content material traits to pick out the optimum mixture in your particular use case.

Conclusion

On this publish, we confirmed how Ring constructed a production-ready, multi-locale RAG-based help chatbot utilizing Amazon Bedrock Information Bases. The structure combines automated content material ingestion, systematic each day analysis utilizing an LLM-as-judge method, and metadata-driven content material concentrating on to attain a 21% discount in infrastructure and operational value per extra locale, whereas sustaining constant buyer experiences throughout 10 worldwide Areas.

Past the core RAG structure, we coated key design issues for manufacturing deployments: vector retailer choice, versioning methods, multi-Area deployment for catastrophe restoration, Cross-Area Inference for scaling basis mannequin throughput, embedding mannequin choice and chunking methods. These patterns apply broadly to any group constructing multi-locale or high-availability RAG programs on AWS.Ring continues to evolve their chatbot structure towards an agentic mannequin with dynamic agent choice and integration of a number of specialised brokers. This agentic method will enable Ring to route buyer inquiries to specialised brokers for gadget troubleshooting, order administration, and product suggestions, demonstrating the extensibility of RAG-based help programs constructed on Amazon Bedrock.

To study extra about Amazon Bedrock Information Bases, go to the Amazon Bedrock documentation.

Concerning the authors