@onioneko/kiki-llm-bedrock

Kiki · Middleware framework for LLM agents

npm install @onioneko/kiki-llm-bedrock

AWS Bedrock LLM provider as middleware. Calls the Bedrock Converse API and pushes the assistant response onto ctx.messages, then calls next() to allow response-processing middleware downstream. Works with any Bedrock-hosted model (Claude, Llama, Mistral, etc.).

Usage

import { compose } from "@onioneko/kiki-core";
import { createBedrockMiddleware } from "@onioneko/kiki-llm-bedrock";

const chain = compose([
  createBedrockMiddleware({ model: "anthropic.claude-sonnet-4-6" }),
]);

const ctx = {
  messages: [{ role: "user", content: "What is 2 + 2?" }],
  system: "Be concise.",
};
await chain(ctx);
// ctx.messages[1] is the assistant response

Configuration

model is required. Other fields are optional with sensible defaults. Authentication uses the standard AWS credential chain (env vars, IAM role, SSO, etc.).

createBedrockMiddleware({
  model: "anthropic.claude-sonnet-4-6", // required
  region: "us-east-1",     // or AWS_REGION env var
  maxTokens: 16384,        // default: 16384
  client: bedrockClient,   // optional: injected BedrockRuntimeClient (for testing)
});

Behavior

• Non-terminal — calls next() after pushing the response, allowing response-processing middleware to be composed after the LLM.
• Converse API — uses the unified ConverseCommand, not the low-level InvokeModelCommand. Model-agnostic.
• Multi-modal — handles both string and ContentBlock[] message content. Images are sent as bytes (base64-decoded).
• No URL images — Bedrock Converse only supports bytes and S3 locations, not direct URLs. URL-type images are skipped.
• AWS credentials — uses the standard AWS credential chain. No API key config needed.
• Config precedence — explicit config > defaults.

Exports

• createBedrockMiddleware(config) — Factory function returning a Middleware
• BedrockConfig — Configuration type
• Re-exports from @onioneko/kiki-types: Message, ConversationContext, ContentBlock, TextBlock, ImageBlock