ADR-037: AWS Bedrock and S3 Vectors Integration

Status

Accepted

Context

The CLI currently supports multiple LLM providers via Quarkus LangChain4j:

Ollama (local, default for chat and embeddings)
Anthropic Claude (cloud chat)
OpenAI (optional cloud chat)

For vector storage, we use PGVector (PostgreSQL with pgvector extension).

Users have requested AWS integration for:

AWS Bedrock - Managed LLM service with access to Claude, Titan, Llama, and other models
Amazon S3 Vectors - New cost-effective vector storage (up to 90% savings vs traditional vector DBs)

Decision

AWS Bedrock Integration (Implemented)

Add quarkus-langchain4j-bedrock extension to support AWS Bedrock as an LLM provider:

Supported Models:

Amazon Titan (text, embedding)
Anthropic Claude (via Bedrock)
Meta Llama
Amazon Nova
Cohere, AI21, Mistral (regional availability varies)

Configuration Pattern:

# Provider selection
quarkus.langchain4j.chat-model.provider=bedrock
quarkus.langchain4j.embedding-model.provider=bedrock

# Bedrock configuration
quarkus.langchain4j.bedrock.region=${AWS_REGION:us-east-1}
quarkus.langchain4j.bedrock.chat-model.model-id=anthropic.claude-3-sonnet-20240229-v1:0
quarkus.langchain4j.bedrock.embedding-model.model-id=amazon.titan-embed-text-v2:0
quarkus.langchain4j.bedrock.timeout=120s

# AWS credentials (use default credential chain)
# Environment variables: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY
# Or: ~/.aws/credentials profile, IAM role, etc.

Benefits:

Enterprise-grade managed service
No infrastructure to maintain
Pay-per-use pricing
Access to latest Claude models without separate API key
VPC endpoints for private connectivity

Amazon S3 Vectors (Not Yet Implemented)

Status: Deferred pending LangChain4j support

Amazon S3 Vectors (GA July 2025) provides native vector storage in S3:

Up to 90% cost reduction vs traditional vector DBs
2 billion vectors per index
Sub-second query latency
Strong consistency

Current Limitation: LangChain4j (Java) does not have native S3 Vectors support. Only Python LangChain has AmazonS3Vectors integration via langchain_aws.vectorstores.s3_vectors.

Options Considered:

Custom EmbeddingStore implementation - High effort, maintenance burden
Wait for LangChain4j support - Preferred, community-driven
Use Amazon OpenSearch Serverless - Supported via existing LangChain4j OpenSearch module

Recommendation: Continue using PGVector for now. Monitor LangChain4j for S3 Vectors support. Consider OpenSearch Serverless for AWS-native deployments requiring managed vector search.

Implementation

Dependencies Added

&lt;!-- AWS Bedrock LLM/Embedding support --&gt;
&lt;dependency&gt;
    &lt;groupId&gt;io.quarkiverse.langchain4j&lt;/groupId&gt;
    &lt;artifactId&gt;quarkus-langchain4j-bedrock&lt;/artifactId&gt;
    &lt;version&gt;${quarkus-langchain4j.version}&lt;/version&gt;
&lt;/dependency&gt;

Configuration Properties

New properties in application.properties:

# ==================== AWS Bedrock Configuration ====================

# Bedrock model configuration
quarkus.langchain4j.bedrock.region=${AWS_REGION:us-east-1}
quarkus.langchain4j.bedrock.timeout=120s

# Bedrock chat model (Claude 3 Sonnet via Bedrock)
quarkus.langchain4j.bedrock.chat-model.model-id=anthropic.claude-3-sonnet-20240229-v1:0
quarkus.langchain4j.bedrock.chat-model.max-tokens=4096
quarkus.langchain4j.bedrock.chat-model.log-requests=false
quarkus.langchain4j.bedrock.chat-model.log-responses=false

# Bedrock embedding model (Amazon Titan)
quarkus.langchain4j.bedrock.embedding-model.model-id=amazon.titan-embed-text-v2:0

Usage

To use Bedrock instead of Ollama:

# Set provider to bedrock
export LANGCHAIN4J_CHAT_MODEL_PROVIDER=bedrock
export LANGCHAIN4J_EMBEDDING_MODEL_PROVIDER=bedrock

# AWS credentials (one of these methods)
export AWS_ACCESS_KEY_ID=your-key
export AWS_SECRET_ACCESS_KEY=your-secret
export AWS_REGION=us-east-1

# Or use ~/.aws/credentials profile
export AWS_PROFILE=your-profile

# Run CLI
java -jar idempiere-hub-runner.jar ask &quot;What is iDempiere?&quot;

Consequences

Positive

Enterprise AWS customers can use managed LLM service
No need for separate Anthropic API key when using Claude via Bedrock
Cost optimization through AWS billing integration
IAM-based access control for LLM usage
Foundation for future S3 Vectors integration

Negative

AWS-specific configuration adds complexity for non-AWS users
Bedrock model availability varies by region
Embedding dimension may differ between models (Titan: 1024, nomic-embed-text: 768)

Neutral

S3 Vectors integration deferred until LangChain4j support is available
PGVector remains the recommended vector store for now