lex-bedrock

Legion Extension that connects LegionIO to AWS Bedrock for foundation model inference.

Purpose

Wraps the AWS Bedrock SDK as named runners consumable by any LegionIO task chain. Provides the Converse API (multi-turn chat with streaming and extended thinking), raw InvokeModel (direct model invocation), foundation model listing, token counting, and inference profile management. Uses the AWS SDK directly — no Faraday. Use this extension when you need direct access to the Bedrock API surface within the LEX runner/actor lifecycle. For simple chat/embed workflows, consider legion-llm instead.

Installation

Add to your Gemfile:

gem 'lex-bedrock'

Configuration

AWS credentials are required:

• access_key_id — AWS access key ID
• secret_access_key — AWS secret access key
• region — AWS region (default: us-east-2)
• session_token — optional, for assumed roles

Usage

Standalone Client

client = Legion::Extensions::Bedrock::Client.new(
  access_key_id: 'AKIA...',
  secret_access_key: 'secret',
  region: 'us-east-2'
)

# List foundation models
result = client.list
# => { models: [...] }

# Converse API
result = client.create(
  model_id: 'us.anthropic.claude-sonnet-4-6',
  messages: [{ role: 'user', content: [{ text: 'Hello' }] }]
)
# => { result: ..., usage: ..., stop_reason: 'end_turn' }

# Streaming Converse
client.create_stream(
  model_id: 'us.anthropic.claude-sonnet-4-6',
  messages: [{ role: 'user', content: [{ text: 'Tell me a story.' }] }]
) do |event|
  case event[:type]
  when :delta then print event[:text]
  when :stop  then puts "\nStop: #{event[:stop_reason]}"
  when :usage then puts event[:usage].inspect
  end
end

# Extended thinking
result = client.create_with_thinking(
  model_id: 'us.anthropic.claude-sonnet-4-6',
  messages: [{ role: 'user', content: [{ text: 'Solve this step by step.' }] }],
  budget_tokens: 8192
)

# Token counting
result = client.count_tokens(
  model_id: 'us.anthropic.claude-sonnet-4-6',
  messages: [{ role: 'user', content: [{ text: 'Hello' }] }]
)
# => { input_token_count: 10 }

# Raw InvokeModel
result = client.invoke_model(
  model_id: 'anthropic.claude-3-5-sonnet-20241022-v2:0',
  body: { prompt: 'Hello', max_tokens: 256 }
)

# Resolve model ID via ModelRegistry
resolved = Legion::Extensions::Bedrock::Helpers::ModelRegistry.resolve('claude-sonnet-4-6')
# => 'us.anthropic.claude-sonnet-4-6'

Runner Modules

include Legion::Extensions::Bedrock::Runners::Models
include Legion::Extensions::Bedrock::Runners::Converse
include Legion::Extensions::Bedrock::Runners::Invoke
include Legion::Extensions::Bedrock::Runners::Tokens
include Legion::Extensions::Bedrock::Runners::Profiles

API Coverage

Runner	Methods
Converse	create, create_stream, create_with_thinking
Invoke	invoke_model
Models	list, get
Tokens	count_tokens
Profiles	list_inference_profiles, get_inference_profile, resolve_profile_id

Model Registry

Helpers::ModelRegistry maps canonical model names to Bedrock cross-region inference profile IDs:

Legion::Extensions::Bedrock::Helpers::ModelRegistry.resolve('claude-sonnet-4-6')
# => 'us.anthropic.claude-sonnet-4-6'

Legion::Extensions::Bedrock::Helpers::ModelRegistry.all
# => { 'claude-3-5-haiku-20241022' => 'us.anthropic.claude-3-5-haiku-20241022-v1:0', ... }

Extended Thinking

Helpers::Thinking provides field builders for Anthropic extended thinking on Bedrock:

# Fixed budget
fields = Legion::Extensions::Bedrock::Helpers::Thinking.build_thinking_fields(budget_tokens: 8192)

# Adaptive thinking
fields = Legion::Extensions::Bedrock::Helpers::Thinking.build_thinking_fields(adaptive: true)

Constants: THINKING_BETA = 'interleaved-thinking-2025-05-14', CONTEXT_1M_BETA = 'context-1m-2025-08-07', TOOL_SEARCH_BETA = 'tool-search-tool-2025-10-19'.

Related

• lex-claude — Direct Anthropic API client (bypasses Bedrock)
• legion-llm — High-level LLM interface including Bedrock via ruby_llm
• extensions-ai/CLAUDE.md — Architecture patterns shared across all AI extensions

Version

0.2.1

License

MIT