Next.jsUnverified

@ai-sdk/google Gemini 2.5 Flash Function Calling Fails — Zod Thought Part Validation Error

Production breaking: Vercel AI SDK @ai-sdk/google suddenly fails for all Gemini 2.5 Flash users. Function calling streams throw AI_TypeValidationError because Gemini introduced 'thought' parts (reasoning tokens) in API responses that the SDK Zod schema did not recognize — it only accepted text/functionCall/inlineData. 32 GitHub reactions, 22 comments, fix confirmed by Vercel team and shipped in @ai-sdk/google@1.2.19. Three resolution paths: upgrade, disable thinking, or downgrade model. Affects all Next.js / Vercel AI SDK apps using Gemini 2.5 Flash with tools.

Symptoms

Streaming function/tool calls fail with: AI_TypeValidationError: Type validation failed: candidates[0].content.parts[0] — expects text/functionCall/inlineData but received none
The failing part contains 'thought: true' and 'thoughtSignature' — Gemini 2.5 Flash thinking/reasoning tokens that the SDK cannot parse
All function calling breaks, not intermittent — error is deterministic when Gemini decides to think before calling a tool
Downgrading to Gemini 2.0 Flash immediately resolves the issue (2.0 does not emit thought parts)
Error appears suddenly without code changes — triggered by Gemini 2.5 Flash API behavior change on Google's side

Possible causes

Gemini 2.5 Flash API silently introduced 'thought' parts in streaming responses — a new content part type for the model's internal reasoning before function calls
The AI SDK @ai-sdk/google response parser uses Zod union: content.parts[] expects text | functionCall | inlineData. Thought is not in this union — Zod rejects the payload
This is a server-side API change by Google, not an SDK regression — the model began emitting thought parts for certain queries, exposing the SDK's incomplete schema
Trigger depends on query complexity: Gemini 2.5 Flash decides internally whether to think before a function call, explaining why users saw the error 'suddenly' without deploying changes

Solutions

Roll back to Gemini 2.0 Flash model (fastest emergency fix)

risk: lowgithubpending_review

Switch model identifier from gemini-2.5-flash-preview-04-17 to gemini-2.0-flash-001. Gemini 2.0 does not emit thought parts, avoiding the issue entirely. Fastest production recovery if you cannot wait for SDK upgrade or disable thinking.

Change model string from 'gemini-2.5-flash-preview-04-17' to 'gemini-2.0-flash-001'
Redeploy — no other code changes needed
Monitor response quality: 2.0 has lower reasoning capability than 2.5 Flash

Commands

grep -r 'gemini-2.5-flash' src/

# Find all occurrences, replace with: gemini-2.0-flash-001

pnpm build && pnpm start

# Verify no AI_TypeValidationError in production logs

Config examples

// Before (broken with @ai-sdk/google < 1.2.19):
const model = google('gemini-2.5-flash-preview-04-17');

// After (working immediately):
const model = google('gemini-2.0-flash-001');

Risks

Gemini 2.0 Flash has lower reasoning quality than 2.5 Flash
Temporary rollback — plan to upgrade @ai-sdk/google and switch back to 2.5

Verification

grep -r 'gemini-2.5-flash' src/ # should return no results after replacement
Deploy and run function calling test
# Expected: zero AI_TypeValidationError, streaming functions complete

✓ 0 verified✕ 0 failed

Disable Gemini thinking/reasoning (workaround without upgrade)

risk: lowgithubpending_review

If you cannot immediately upgrade @ai-sdk/google, suppress thought parts by setting thinkingBudget: 0 in providerOptions. This tells Gemini not to emit reasoning tokens, sidestepping the Zod validation error.

Add providerOptions.google.thinkingConfig.thinkingBudget: 0 to your streamText/generateText call
This prevents thought parts from appearing in responses entirely
Note: may reduce reasoning quality for complex tool-use scenarios — treat as temporary

Commands

node -e "
import { google } from '@ai-sdk/google';
import { streamText } from 'ai';
const r = await streamText({
  model: google('gemini-2.5-flash-preview-04-17'),
  providerOptions: { google: { thinkingConfig: { thinkingBudget: 0 } } },
  tools: { test: { description: 'test', parameters: {} } },
  messages: [{ role: 'user', content: 'test' }]
});
console.log('OK — no AI_TypeValidationError');
"

Config examples

import { google } from '@ai-sdk/google';
import { streamText } from 'ai';

const result = streamText({
  model: google('gemini-2.5-flash-preview-04-17'),
  providerOptions: {
    google: {
      thinkingConfig: {
        thinkingBudget: 0,  // disables thought parts
      },
    },
  },
  tools: { /* your tools */ },
  messages,
});

Risks

Disabling thinking may reduce reasoning quality for complex tool-use scenarios
This is a workaround, not a fix — upgrade @ai-sdk/google when possible

Verification

Add thinkingConfig: { thinkingBudget: 0 } to your providerOptions
Run your function calling test with Gemini 2.5 Flash
# Expected: streaming completes without AI_TypeValidationError

✓ 0 verified✕ 0 failed

Upgrade @ai-sdk/google to 1.2.19+ (permanent fix)

risk: lowgithubpending_review

Vercel team shipped a fix in @ai-sdk/google@1.2.19 that adds thought part support to the response Zod schema. This is the recommended solution — confirmed working by multiple users in the issue thread and by Vercel staff (@philschmid).

Upgrade: pnpm add @ai-sdk/google@latest (or npm install @ai-sdk/google@latest)
Verify: pnpm list @ai-sdk/google should show >= 1.2.19
Redeploy and test with Gemini 2.5 Flash + function calling

Commands

pnpm add @ai-sdk/google@latest

pnpm list @ai-sdk/google | grep @ai-sdk/google

# Expected output: @ai-sdk/google 1.2.19 (or higher)

Config examples

# package.json after upgrade:
{
  "dependencies": {
    "@ai-sdk/google": "^1.2.19",
    "ai": "^4.x"
  }
}

Verification

pnpm list @ai-sdk/google | grep @ai-sdk/google
# Expected: @ai-sdk/google 1.2.19 or higher
Test: streamText({ model: google('gemini-2.5-flash-preview-04-17'), tools: {...} })
# Expected: streaming function calls complete without AI_TypeValidationError

✓ 0 verified✕ 0 failed

Agent JSON

Canonical machine-readable representation of this issue:

{
  "issue_id": "6c4cb49a-ab2c-4410-b07e-31944a227ff5",
  "slug": "ai-sdk-google-gemini-2-5-flash-function-calling-fails-zod-thought-part-validation-error-8vpu3k",
  "verification_status": "unverified",
  "canonical_json": "https://codekb.dev/v1/issues/ai-sdk-google-gemini-2-5-flash-function-calling-fails-zod-thought-part-validation-error-8vpu3k"
}