Add AWS Bedrock provider support

[moncapitaine]

Overview

Add support for AWS Bedrock as an AI provider, enabling users to access multiple foundation models (Anthropic, Meta, Cohere, Amazon) through a unified AWS-native platform.

Why AWS Bedrock?

• Multi-provider platform: Access Anthropic Claude, Meta Llama, Cohere, Amazon Titan, etc.
• AWS-native: Seamless integration with AWS services (S3, Lambda, SageMaker)
• Enterprise AWS customers: Organizations already using AWS infrastructure
• Regional compliance: AWS regional deployments for data residency
• Unified billing: Single AWS bill for all models

Capabilities

• ✅ Chat Completion: Claude (Anthropic), Llama (Meta), Command (Cohere), Titan (Amazon)
• ✅ Embedding: Titan Embeddings, Cohere Embed
• ✅ Vision: Claude 3 models (via Bedrock)
• ⚠️ Function Calling: Model-dependent (Claude supports it)

Implementation Checklist

Backend

• Create AWS Bedrock provider client in ai-service-client package
• Implement AWS SDK v3 authentication (IAM, credentials)
• Implement chat completion (InvokeModel API)
• Implement streaming support (InvokeModelWithResponseStream)
• Implement embedding generation
• Handle model-specific request/response formats
• Add model capability detection
• Add to provider cache system
• Add connection testing endpoint
• Support cross-region model access

Database

• Add aws_bedrock to provider enum (if needed)
• Update provider configuration schema
• Support AWS-specific fields:
  • region (AWS region)
  • accessKeyId (optional, can use IAM role)
  • secretAccessKey (optional)

GraphQL API

• Add AWS Bedrock provider to AiServiceProvider mutations
• Model discovery from Bedrock ListFoundationModels API
• Support region selection
• Handle IAM role vs access key authentication

Frontend

• Add AWS Bedrock provider UI in /admin/ai-services
• Configuration form:
  • AWS region
  • Access credentials (or IAM role)
  • Model selection
• Connection testing
• Model selection filtered by Bedrock
• Display model providers (Anthropic, Meta, etc.)

Documentation

• Update /docs/admin/ai-models with AWS Bedrock
• Update /docs/admin/ai-services with Bedrock configuration
• Add AWS Bedrock setup guide
• Document IAM role setup
• Document regional model availability
• Explain multi-provider access

Testing

• Unit tests for Bedrock client
• Integration tests for chat completion (multiple models)
• Integration tests for embeddings
• E2E tests for provider configuration
• Test IAM authentication
• Test cross-region access

API Details

SDK: AWS SDK for JavaScript v3 (@aws-sdk/client-bedrock-runtime)
Authentication: IAM roles, access keys, or STS temporary credentials
API: InvokeModel, InvokeModelWithResponseStream, ListFoundationModels
Region-based: Models available in specific regions (us-east-1, eu-west-1, etc.)

Model ID Examples:

• anthropic.claude-3-5-sonnet-20241022-v2:0
• meta.llama3-2-90b-instruct-v1:0
• cohere.command-r-plus-v1:0
• amazon.titan-embed-text-v2:0

Challenges

1. Complex authentication: IAM roles, cross-account access
2. Model-specific formats: Different request/response formats per provider
3. Regional availability: Not all models in all regions
4. AWS SDK dependency: Large SDK, adds complexity
5. Cost tracking: Harder to attribute costs per model

Resources

• AWS Bedrock Documentation
• Supported Models
• API Reference
• IAM Permissions

Related

• Part of multi-provider support initiative
• Enables AWS-native deployments
• Provides unified access to multiple model providers

Priority

P3-low - Valuable for AWS-heavy organizations, but adds significant complexity. Most users can access Anthropic/Meta/Cohere directly.