AWS Transformのモデル移行機能を実際に試してみた

Discover → Clarify → Map → Estimate → Recommend

curl -fsSL https://transform-cli.awsstatic.com/install.sh | bash
atx --version
# v3.2.0

eval "$(aws configure export-credentials --profile kiro-aws --format env)"
unset AWS_PROFILE
export AWS_REGION=ap-northeast-1

aws sts get-caller-identity

$ atx custom def list

AWS/GenAI-to-Bedrock-Migration-Assessment

import { GoogleGenAI, type Content } from "@google/genai";

function selectModel(userMessage: string): string {
  const isSimpleQuery = userMessage.length < 100 && !userMessage.includes("?");
  return isSimpleQuery ? "gemini-2.5-flash-lite" : "gemini-2.5-flash";
}

export class GeminiChatbot {
  private ai: GoogleGenAI;
  private history: Content[] = [];

  constructor() {
    this.ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY ?? "" });
  }

  async chat(userMessage: string): Promise<string> {
    const modelName = selectModel(userMessage);
    this.history.push({ role: "user", parts: [{ text: userMessage }] });

    const response = await this.ai.models.generateContent({
      model: modelName,
      contents: this.history,
      config: { maxOutputTokens: 1024, temperature: 0.7 },
    });

    const assistantMessage = response.text ?? "";
    this.history.push({ role: "model", parts: [{ text: assistantMessage }] });
    return assistantMessage;
  }
}

cd ~/work/aws-transform-m2m

atx custom def exec \
  --code-repository-path ./gemini-flash-chatbot \
  --transformation-name AWS/GenAI-to-Bedrock-Migration-Assessment \
  --trust-all-tools

$ atx custom def exec \
    --code-repository-path ./gemini-flash-chatbot \
    --transformation-name AWS/GenAI-to-Bedrock-Migration-Assessment \
    --trust-all-tools

📝 Conversation log: .../logs/2026-06-17T05-16-42-conversation.log

{
  "summary": {
    "overall_confidence": 0.98,
    "confidence_level": "very_high",
    "total_models_detected": 2,
    "languages_found": ["typescript"],
    "ai_source": "gemini"
  },
  "models": [
    {
      "model_id": "gemini-2.5-flash",
      "detected_via": ["code"],
      "evidence": [
        {
          "source": "code",
          "file": "src/chatbot.ts",
          "line": 19,
          "pattern": "return isSimpleQuery ? \"gemini-2.5-flash-lite\" : \"gemini-2.5-flash\""
        }
      ],
      "usage_context": "Primary chatbot model for complex queries and multi-turn conversation"
    },
    {
      "model_id": "gemini-2.5-flash-lite",
      "usage_context": "Cost-optimized model for simple/short queries via tiered routing"
    }
  ],
  "integration": {
    "primary_sdk": "@google/genai",
    "sdk_version": "^1.3.0",
    "pattern": "direct_sdk",
    "capabilities_summary": {
      "text_generation": true,
      "streaming": false,
      "function_calling": false
    }
  }
}

ATX: 5 quick questions (reply like "A, E, A, B, A"):

1. Framework:    A) Direct API *(detected)* | B) Router | C) LangChain | ...
2. Priority:     A) Quality | B) Speed | C) Cost | E) Balanced | F) Don't know
3. Monthly spend: A) <$500 | B) $500–$2K | C) $2K–$10K | D) >$10K
4. Cross-cloud:  A) Latency critical | B) Acceptable | C) Egress concern | D) Test first
5. Model:        A) Gemini Flash *(detected)* | B) Gemini Pro | J) Other

> A, C, A, B, A   ← Direct API / Cost 優先 / <$500 / Latency acceptable / Gemini Flash

ATX: Last section — 5 questions about your technical requirements:

Q6. Input types:   A) Text only | B) Vision required | C) Audio/Video
Q7. Monthly volume: A) < 1M tokens | B) 1–10M | C) 10–100M | D) > 100M
Q8. Response speed: A) Critical (<500ms) | B) Important (<2s) | C) Flexible
Q9. Complexity:    A) Simple | B) Moderate | C) Complex
Q10. Features:     A) Function calling | ... | J) None

> A, A, B, B, J

{
  "ai_constraints": {
    "ai_framework":     { "value": ["direct"], "chosen_by": "extracted" },
    "ai_priority":      { "value": "cost",     "chosen_by": "user" },
    "ai_monthly_spend": { "value": "<$500",    "chosen_by": "user" },
    "cross_cloud":      { "value": "latency-acceptable", "chosen_by": "user" },
    "ai_model_baseline":{ "value": "claude-haiku-4-5",  "chosen_by": "user" },
    "ai_vision":        { "value": "text-only", "chosen_by": "user" },
    "ai_token_volume":  { "value": "low",       "chosen_by": "user" },
    "ai_latency":       { "value": "important", "chosen_by": "user" },
    "ai_complexity":    { "value": "moderate",  "chosen_by": "user" }
  }
}

ATX: Design Summary:

| Source Model          | Bedrock Model      | Price (in/out per 1M) | Savings     | Assessment     |
|-----------------------|--------------------|-----------------------|-------------|----------------|
| gemini-2.5-flash      | Nova Lite (v1:0)   | $0.06 / $0.24         | 88% cheaper | Strong migrate |
| gemini-2.5-flash-lite | Nova Micro (v1:0)  | $0.035 / $0.14        | 53% cheaper | Strong migrate |

These models work as a cascade/routing pattern (flash-lite for simple queries, flash for complex).
Test the Bedrock replacement pair together — validate that Nova Micro's quality is acceptable
for simple queries before testing Nova Lite for complex queries.

ATX: Cost Estimation Summary (cached pricing, ±15-25% accuracy):

| Model Option                         | Monthly Cost | vs Current ($7.74/mo) |
|--------------------------------------|--------------|------------------------|
| Nova Lite + Nova Micro (recommended) | $1.10        | -86%                   |
| Nova Micro only                      | $0.77        | -90%                   |
| Llama 4 Scout                        | $3.66        | -53%                   |
| Llama 4 Maverick                     | $5.32        | -31%                   |
| Claude Haiku 4.5                     | $26.00       | +236%                  |

Recommended: Nova Lite + Nova Micro = $1.10/mo (saves $6.64/mo, $79.68/yr)

Optimization: Nova Micro only (if quality acceptable) → $0.77/mo (-90%)

{
  "migration_plan": {
    "total_weeks": 2,
    "approach": "direct_sdk_swap",
    "phases": [
      {
        "name": "Setup & Development (Week 1)",
        "activities": [
          "Enable Bedrock model access (Nova Lite, Nova Micro) in us-east-1",
          "Create IAM role with bedrock:InvokeModel permission",
          "Install @aws-sdk/client-bedrock-runtime",
          "Develop Bedrock provider adapter replacing @google/genai calls",
          "Implement AI_PROVIDER feature flag for gradual rollout",
          "Preserve tiered routing logic (simple → Nova Micro, complex → Nova Lite)"
        ]
      },
      {
        "name": "Testing & Rollout (Week 2)",
        "activities": [
          "Deploy to staging with AI_PROVIDER=bedrock",
          "A/B comparison: 100+ representative prompts",
          "Gradual rollout: 10% → 50% → 100%",
          "Monitor 48h at 100%"
        ]
      }
    ]
  }
}

import { GoogleGenAI } from '@google/genai';

const ai = new GoogleGenAI({ apiKey: GEMINI_API_KEY });
const response = await ai.models.generateContent({
  model: modelName,             // "gemini-2.5-flash" or "gemini-2.5-flash-lite"
  contents: history,            // Content[] 形式
  config: {
    systemInstruction,
    maxOutputTokens: 1024,
    temperature: 0.7
  }
});
const text = response.text;

import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';

const client = new BedrockRuntimeClient({ region: 'us-east-1' });
const response = await client.send(new ConverseCommand({
  modelId: modelName,                   // "amazon.nova-lite-v1:0" or "amazon.nova-micro-v1:0"
  messages: messages,                   // {role, content: [{text}]}[] 形式
  system: [{ text: systemInstruction }],
  inferenceConfig: {
    maxTokens: 1024,
    temperature: 0.7
  }
}));
const text = response.output.message.content[0].text;

.migration/0617-0517/ai-migration/
├── provider_adapter.js    # gemini/bedrock/shadow 3モードの Provider Adapter
├── setup_bedrock.sh       # Bedrock セットアップ（dry-run 対応）
└── test_comparison.py     # Gemini vs Bedrock A/B テストハーネス

{
  "mechanism": "feature_flag",
  "flag_name": "AI_PROVIDER",
  "default_value": "gemini",
  "rollback_time": "instant (env var change)",
  "triggers": [
    "Response quality below 80% of Gemini baseline",
    "P95 latency > 2x Gemini baseline for 5 minutes",
    "Error rate > 1% for 5 consecutive minutes"
  ]
}

Branch: atx-result-staging-20260617_051633_d4197220
Commit: Step 1: Complete Gemini-to-Bedrock migration assessment.
        Phases: Discover (detected gemini-2.5-flash + flash-lite),
        Clarify (AI-only, cost priority),
        Design (Nova Lite + Nova Micro, 86% savings),
        Estimate ($7.74/mo → $1.10/mo),
        Generate (setup script, provider adapter, test harness, migration guide)