Migrate @google/generative-ai to @google/genai (JS/TS)

What changed

• new GoogleGenerativeAI(key) becomes new GoogleGenAI({ apiKey }).
• genAI.getGenerativeModel({ model }) is gone. Pass model to each call on ai.models.
• model.generateContent(prompt) becomes ai.models.generateContent({ model, contents }), and the result is the response itself, not { response }.
• response.text() is now a property: response.text.
• Streams are iterated directly: for await (const chunk of stream) instead of result.stream.
• generationConfig and safetySettings move into a per-request config object.
• GoogleAIFileManager and GoogleAICacheManager from @google/generative-ai/server are replaced by ai.files and ai.caches.

Migration steps

1. Swap the package

   npm uninstall @google/generative-ai
   npm install @google/genai

   Copy

   The README lists Node.js 20 or later as a prerequisite and warns that SDK version 3.0.0 will require Node.js 22. It also says not to ship API keys in browser code; keep calls on the server.

2. Create a client and call ai.models

   Before

   import { GoogleGenerativeAI } from "@google/generative-ai";
   
   const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY);
   const model = genAI.getGenerativeModel({ model: "gemini-flash-latest" });
   
   const result = await model.generateContent("Tell me a story in 300 words");
   console.log(result.response.text());

   Copy

   After

   import { GoogleGenAI } from "@google/genai";
   
   const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });
   
   const response = await ai.models.generateContent({
     model: "gemini-flash-latest",
     contents: "Tell me a story in 300 words",
   });
   console.log(response.text);

   Copy

   Google's code generation notes also list the type renames: GenerateContentResult becomes GenerateContentResponse and GenerateContentRequest becomes GenerateContentParameters.

3. Move generationConfig into config

   Before

   const model = genAI.getGenerativeModel({
     model: "gemini-flash-latest",
     generationConfig: { stopSequences: ["x"], maxOutputTokens: 20, temperature: 1.0 },
   });
   const result = await model.generateContent("Tell me a story about a magic backpack.");

   Copy

   After

   const response = await ai.models.generateContent({
     model: "gemini-flash-latest",
     contents: "Tell me a story about a magic backpack.",
     config: {
       stopSequences: ["x"],
       maxOutputTokens: 20,
       temperature: 1.0,
       systemInstruction: "You are a storyteller for young children",
     },
   });

   Copy

   Settings that used to live on the model object are now per request. safetySettings goes in the same config, with string categories such as "HARM_CATEGORY_HARASSMENT". If several call sites shared one configured model, keep a shared config object and spread it into each call.

4. Iterate streams directly

   Before

   const result = await model.generateContentStream(prompt);
   for await (const chunk of result.stream) {
     process.stdout.write(chunk.text());
   }

   Copy

   After

   const stream = await ai.models.generateContentStream({
     model: "gemini-flash-latest",
     contents: prompt,
   });
   for await (const chunk of stream) {
     process.stdout.write(chunk.text ?? "");
   }

   Copy

5. Create chats from ai.chats

   Before

   const chat = model.startChat({
     history: [
       { role: "user", parts: [{ text: "Hello" }] },
       { role: "model", parts: [{ text: "Great to meet you. What would you like to know?" }] },
     ],
   });
   const result = await chat.sendMessage("I have 2 dogs in my house.");
   console.log(result.response.text());

   Copy

   After

   const chat = ai.chats.create({
     model: "gemini-flash-latest",
     history: [
       { role: "user", parts: [{ text: "Hello" }] },
       { role: "model", parts: [{ text: "Great to meet you. What would you like to know?" }] },
     ],
   });
   const response = await chat.sendMessage({ message: "I have 2 dogs in my house." });
   console.log(response.text);

   Copy

   sendMessage takes an object with a message field instead of a bare string. For streaming chats use chat.sendMessageStream({ message }) and iterate the result with for await.

6. Replace the server helpers and other model methods

   • new GoogleAIFileManager(key).uploadFile(...) to ai.files.upload({ file })
   • new GoogleAICacheManager(key).create(...) to ai.caches.create(...)
   • model.countTokens(prompt) to ai.models.countTokens({ model, contents: prompt }); totalTokens is still on the result
   • model.embedContent(text) to ai.models.embedContent({ model: "gemini-embedding-001", contents: text }), which returns embeddings (plural)
   • result.response.usageMetadata to response.usageMetadata

7. Catch ApiError

   import { ApiError } from "@google/genai";
   
   try {
     await ai.models.generateContent({ model: "non-existent-model", contents: "Hi" });
   } catch (e) {
     if (e instanceof ApiError) console.error(e.status, e.message);
     else throw e;
   }

   Copy

   ApiError extends Error and adds status, the HTTP code.

How to find every affected call

Use ast-grep to match the syntax tree rather than text, so calls split across lines are found. $GENAI and $MODEL match whatever you named the instances:

sg run --pattern 'new GoogleGenerativeAI($$$ARGS)' --lang ts
sg run --pattern '$GENAI.getGenerativeModel($$$ARGS)' --lang ts
sg run --pattern '$RESULT.response.text()' --lang ts
sg run --pattern '$RESULT.stream' --lang ts
sg run --pattern '$MODEL.startChat($$$ARGS)' --lang ts
sg run --pattern 'new GoogleAIFileManager($$$ARGS)' --lang ts

Repeat with --lang tsx and --lang js. $RESULT.stream also matches unrelated objects with a stream property, so review each hit. chunk.text() inside old stream loops is a separate search: sg run --pattern '$CHUNK.text()' --lang ts.

Testing the change

• Run tsc --noEmit. Leftover .text() calls, result.response reads and old type names show up as type errors in TypeScript files.
• Update mocks. A stub of getGenerativeModel that returns { response: { text: () => "..." } } keeps passing while production breaks; return the new shape with text as a string.
• Run one streamed request and one chat in staging, since both changed how you iterate and how you send a message.
• Trigger an error with a bad model name and confirm your handler sees an ApiError with a numeric status.

Automating it

Google does not publish a codemod. It publishes codegen_instructions.md, a list of incorrect and correct patterns meant to be pasted into an AI coding tool so it writes new-SDK code. That helps a model avoid old calls, but nothing checks that the result builds.

Versyka finds the call sites deterministically with ast-grep, rewrites them together with the code that depends on them (the .text() reads, stream loops, mocks), then runs your repository's install, typecheck, build and tests in a disposable sandbox. A pull request is opened only when those pass.