Example of code to invoke Llama 2 LLM on Amazon Bedrock in the Swift programming language

bedrock_llama2.swift

import Foundation
import ClientRuntime

// reduce the verbosity of the AWS SDK
SDKLoggingSystem.initialize(logLevel: .warning)

import AWSBedrock
import AWSBedrockRuntime

// create a Bedrock client and list available models for a provider
let provider = "meta"
print("====== Models available for \(provider)")
let client = try  BedrockClient(region: "us-east-1")
let input = ListFoundationModelsInput(byProvider : provider)
let output = try await client.listFoundationModels(input: input)
print(output.modelSummaries!.map { "\($0.modelName!) : \($0.modelId!)" }.joined(separator: "\n"))
print("======")

// create a bedrock runtime client and invoke a model
let modelId = "meta.llama2-13b-chat-v1"
let runtime = try BedrockRuntimeClient(region: "us-east-1")
let payload = LlamaChatMessage(prompt: "What is the difference between a llama and an alpaca?")
let request = InvokeModelInput(body: try payload.encode(),
                               contentType: "application/json",
                               modelId: modelId)

let response = try await runtime.invokeModel(input: request)
let llamaResponse = try LlamaChatResponse(from: response!.body!)
print(llamaResponse)
print("======")


struct LlamaChatMessage: Encodable {
    let prompt: String
    let maxGenLen: Int = 512
    let topP: Double = 0.9
    let temperature: Double = 0.2
    
    func encode() throws -> Data {
        let encoder = JSONEncoder()
        encoder.keyEncodingStrategy = .convertToSnakeCase
        return try encoder.encode(self)
    }
}

struct LlamaChatResponse: Decodable, CustomStringConvertible {
    
    let generation: String
    let promptTokenCount: Int
    let generationTokenCount : Int
    
    init(from data: Data) throws {
        let decoder = JSONDecoder()
        decoder.keyDecodingStrategy = .convertFromSnakeCase
        self = try decoder.decode(LlamaChatResponse.self, from: data)
    }
    
    var description: String {
        return generation.trim()
    }
}