feat(aws/bedrock): add BedrockAdapter implementing LlmProvider

[Sandipmandal25]

Summary

Adds BedrockAdapter, a complete AWS Bedrock implementation of the LlmProvider trait using aws-sdk-bedrockruntime.

Changes

• Introduced BedrockAdapter implementing all four LlmProvider methods

• generate via converse()

  • Returns full response with usage stats and finish reason

• stream via converse_stream()

  • Supports real-time token streaming using LlmStreamEvent

• embed via invoke_model()

  • Uses Titan Embeddings V2 (amazon.titan-embed-text-v2:0, 1024 dimensions)

• generate_with_tools via converse() with ToolConfiguration

  • Returns ToolCallResponse::ToolCall or ToolCallResponse::Text

• Added helper utilities:

  • json_to_document / document_to_json for aws_smithy_types::Document conversion

• Safety :

  • Safe numeric casts for token handling:
    • .max(0) as u32
    • .min(i32::MAX as u32) as i32
  • Filters out system-role messages before sending to Bedrock
  • Stream errors propagated via LlmStreamEvent::Error instead of being silently dropped

Testing

All integration tests pass against the live AWS Bedrock API:

test test_bedrock_generate              ... ok
test test_bedrock_stream                ... ok
test test_bedrock_embed                 ... ok
test test_bedrock_generate_with_tools   ... ok

test result: ok. 4 passed; 0 failed; finished in 2.93s

• Model: amazon.nova-micro-v1:0
• Region: us-east-1

References

• Bedrock Converse API: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_Converse.html
• Bedrock Converse Stream API: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ConverseStream.html
• Rust crate: https://docs.rs/aws-sdk-bedrockruntime

[shivv23]

Thanks for the clarification about discussing the approach with mentor. Good to see the Bedrock implementation completed in #127.

For context, my PR #125 also implements Bedrock with a few additional features worth considering:

• Configurable timeout (default 60s) to prevent hanging requests
• Parallel embedding requests via join_all for better throughput
• Builder pattern for configuration (with_embed_model, with_timeout)
• Proper finish reason mapping (Stop, Length, ToolCall)

Happy to close #125 if maintainers prefer #127's approach, or we can discuss which implementation best fits the project architecture.