Groq models fail with json_schema and web_search tool schema errors after 1.12.1 update

[pfjarschel]

Description
After updating to the latest Perplexica version, all my Groq-based providers that were working minutes before the upgrade started failing with structured-output–related errors. Downgrading to the 1.11.2 version immediately fixes the problem, using the exact same Groq key and models.

Environment

• Perplexica version: 1.12.1
• Deployment: Docker (docker-compose)
• Provider: Groq
• Models used:
  • openai/gpt-oss-20b (GPT-OSS)
  • moonshotai/kimi-k2-instruct (KIMI K2)
  • meta-llama/llama-4-maverick-17b-128e-instruct (Llama 4 Maverick)

All of these models support json_schema structured outputs in Groq’s docs.

What happens
SearXNG and Perplexica start normally, but on the first query using a Groq model, nothing happens visually (stuck in Brainstorming), but there is this message in the logs:

Error: 400 This model does not support response format `json_schema`. See supported models at https://console.groq.com/docs/structured-outputs#supported-models
    at f.generate (.../chunks/607.js:20:19183)
    at cX.makeStatusError (.../chunks/607.js:27:51395)
    at cX.makeRequest (.../chunks/607.js:27:54864)
    at async M.generateObject (.../chunks/136.js:1:3636)
    at async h (.../chunks/641.js:464:844)
    at async m.searchAsync (.../app/api/chat/route.js:1:12737)

This repeats for each request, even though the models used are listed as supporting json_schema in Groq’s documentation.

Then there is another error, appears to be related to tools:

Error: tool call validation failed: parameters for tool web_search did not match schema: errors: [missing properties: 'type']
    at bf.f [as iterator] (.../chunks/607.js:20:27895)
    at async M.streamText (.../chunks/136.js:1:3089)
    at async i.research (.../chunks/641.js:541:227)
    at async m.searchAsync (.../app/api/chat/route.js:1:13218)
code: 'tool_use_failed'
type: 'invalid_request_error'

What I expected
Queries using supported Groq models with structured outputs to work as before the update.
Tool calls to web_search to respect the current schema and not fail on a missing type property.

What I tried

• Switching between multiple Groq models (GPT-OSS, KIMI K2, Llama 4 Maverick): all fail with the same json_schema error.
• Restarting containers and recreating volumes.
• Downgrading Perplexica to the 1.11.2 version: with the same Groq configuration, everything works again (no errors, normal search and answers).
• Downgrading Perplexica to the 1.12.0 version: Mostly works, but there are some other, possibly unrelated issues (search engines get blocked very quickly, sometimes query fails without any call to the Groq API).
• I haven't tried other cloud providers.

Hypothesis
It looks like the latest Perplexica release changed:

• How response_format: { type: "json_schema", ... } is applied to Groq models, possibly without checking model capabilities or without using the exact model IDs/flags Groq expects.
• The web_search tool schema (adding a required type property) without updating the caller, so tool calls now fail validation with tool_use_failed.

Both issues only appear on the latest version and disappear after downgrading.

Logs
Here are the complete logs for the session:

Starting SearXNG...
Waiting for SearXNG to be ready...
2026-01-12 19:30:12,256 ERROR:searx.engines: loading engine ahmia failed: set engine to inactive!
2026-01-12 19:30:12,385 ERROR:searx.engines: loading engine torch failed: set engine to inactive!
* Serving Flask app 'searx/webapp.py'
* Debug mode: off
2026-01-12 19:30:15,605 ERROR:searx.botdetection: X-Forwarded-For nor X-Real-IP header is set!
SearXNG started successfully (PID: 9)
Starting Perplexica...
  ▲ Next.js 16.0.7
  - Local:         http://a2fe1a74528d:3000
  - Network:       http://a2fe1a74528d:3000

✓ Starting...
Running database migrations...
Skipping already-applied migration: 0000_fuzzy_randall.sql
Skipping already-applied migration: 0001_wise_rockslide.sql
Skipping already-applied migration: 0002_daffy_wrecker.sql
Database migrations completed successfully
✓ Ready in 92ms
Error: 400 This model does not support response format `json_schema`. See supported models at https://console.groq.com/docs/structured-outputs#supported-models
   at f.generate (.next/server/chunks/607.js:20:19183)
   at cX.makeStatusError (.next/server/chunks/607.js:27:51395)
   at cX.makeRequest (.next/server/chunks/607.js:27:54864)
   at async M.generateObject (.next/server/chunks/136.js:1:3636)
   at async h (.next/server/chunks/641.js:464:844)
   at async m.searchAsync (.next/server/app/api/chat/route.js:1:12737) {
 status: 400,
 headers: Headers {
   date: 'Mon, 12 Jan 2026 19:31:47 GMT',
   'content-type': 'application/json',
   'content-length': '230',
   connection: 'keep-alive',
   'cf-ray': '9bcf137bef4f4292-VCP',
   'cache-control': 'private, max-age=0, no-store, no-cache, must-revalidate',
   vary: 'Origin',
   'x-groq-region': 'bra',
   'x-request-id': 'req_01kesv2tbve84sxnd0eks0nejr',
   via: '1.1 google',
   'cf-cache-status': 'DYNAMIC',
   'set-cookie': '__cf_bm=jtP_hJFiKVDsb_KbRrupAnZpLl1Gr9NV1Ne7C5sHbwE-1768246307-1.0.1.1-T6k_qFeXnaHIz5mgtM8Tyq6lSliLwIn2buklH3Is86NmiCMvW_bnVOrcBg_9x34bYMPFUoy.i75C0qtQUlpPh9vX66oW6cvgok
0pLYFPr5U; path=/; expires=Mon, 12-Jan-26 20:01:47 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
   'strict-transport-security': 'max-age=15552000',
   server: 'cloudflare',
   'alt-svc': 'h3=":443"; ma=86400'
 },
 requestID: 'req_01kesv2tbve84sxnd0eks0nejr',
 error: [Object],
 code: undefined,
 param: 'response_format',
 type: 'invalid_request_error'
}
⨯ unhandledRejection:  Error: 400 This model does not support response format `json_schema`. See supported models at https://console.groq.com/docs/structured-outputs#supported-models
   at f.generate (.next/server/chunks/607.js:20:19183)
   at cX.makeStatusError (.next/server/chunks/607.js:27:51395)
   at cX.makeRequest (.next/server/chunks/607.js:27:54864)
   at async M.generateObject (.next/server/chunks/136.js:1:3636)
   at async h (.next/server/chunks/641.js:464:844)
   at async m.searchAsync (.next/server/app/api/chat/route.js:1:12737) {
 status: 400,
 headers: Headers {
   date: 'Mon, 12 Jan 2026 19:31:47 GMT',
   'content-type': 'application/json',
   'content-length': '230',
   connection: 'keep-alive',
   'cf-ray': '9bcf137bef4f4292-VCP',
   'cache-control': 'private, max-age=0, no-store, no-cache, must-revalidate',
   vary: 'Origin',
   'x-groq-region': 'bra',
   'x-request-id': 'req_01kesv2tbve84sxnd0eks0nejr',
   via: '1.1 google',
   'cf-cache-status': 'DYNAMIC',
   'set-cookie': '__cf_bm=jtP_hJFiKVDsb_KbRrupAnZpLl1Gr9NV1Ne7C5sHbwE-1768246307-1.0.1.1-T6k_qFeXnaHIz5mgtM8Tyq6lSliLwIn2buklH3Is86NmiCMvW_bnVOrcBg_9x34bYMPFUoy.i75C0qtQUlpPh9vX66oW6cvgok
0pLYFPr5U; path=/; expires=Mon, 12-Jan-26 20:01:47 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
   'strict-transport-security': 'max-age=15552000',
   server: 'cloudflare',
   'alt-svc': 'h3=":443"; ma=86400'
 },
 requestID: 'req_01kesv2tbve84sxnd0eks0nejr',
 error: [Object],
 code: undefined,
 param: 'response_format',
 type: 'invalid_request_error'
}
Error: 400 This model does not support response format `json_schema`. See supported models at https://console.groq.com/docs/structured-outputs#supported-models
   at f.generate (.next/server/chunks/607.js:20:19183)
   at cX.makeStatusError (.next/server/chunks/607.js:27:51395)
   at cX.makeRequest (.next/server/chunks/607.js:27:54864)
   at async M.generateObject (.next/server/chunks/136.js:1:3636)
   at async h (.next/server/chunks/641.js:464:844)
   at async m.searchAsync (.next/server/app/api/chat/route.js:1:12737) {
 status: 400,
 headers: Headers {
   date: 'Mon, 12 Jan 2026 19:32:55 GMT',
   'content-type': 'application/json',
   'content-length': '230',
   connection: 'keep-alive',
   'cf-ray': '9bcf15268b584293-VCP',
   'cache-control': 'private, max-age=0, no-store, no-cache, must-revalidate',
   vary: 'Origin',
   'x-groq-region': 'bra',
   'x-request-id': 'req_01kesv4x14eptanxe44mfxndqd',
   via: '1.1 google',
   'cf-cache-status': 'DYNAMIC',
   'set-cookie': '__cf_bm=AwxBo2U51m8khHAx3TPmbXRHVrDRBk3G7nLSZBsEFKA-1768246375-1.0.1.1-2ipUeEYn.NixbCYYMofApTbyS02oKlDZX8WpGN3Xl8twVthSMx8lHd1GBZS0bQamwtWOG6yiverHhfzujUw0EQteGSjMzKSC9K
iQ54ioBxE; path=/; expires=Mon, 12-Jan-26 20:02:55 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
   'strict-transport-security': 'max-age=15552000',
   server: 'cloudflare',
   'alt-svc': 'h3=":443"; ma=86400'
 },
 requestID: 'req_01kesv4x14eptanxe44mfxndqd',
 error: [Object],
 code: undefined,
 param: 'response_format',
 type: 'invalid_request_error'
}
⨯ unhandledRejection:  Error: 400 This model does not support response format `json_schema`. See supported models at https://console.groq.com/docs/structured-outputs#supported-models
   at f.generate (.next/server/chunks/607.js:20:19183)
   at cX.makeStatusError (.next/server/chunks/607.js:27:51395)
   at cX.makeRequest (.next/server/chunks/607.js:27:54864)
   at async M.generateObject (.next/server/chunks/136.js:1:3636)
   at async h (.next/server/chunks/641.js:464:844)
   at async m.searchAsync (.next/server/app/api/chat/route.js:1:12737) {
 status: 400,
 headers: Headers {
   date: 'Mon, 12 Jan 2026 19:32:55 GMT',
   'content-type': 'application/json',
   'content-length': '230',
   connection: 'keep-alive',
   'cf-ray': '9bcf15268b584293-VCP',
   'cache-control': 'private, max-age=0, no-store, no-cache, must-revalidate',
   vary: 'Origin',
   'x-groq-region': 'bra',
   'x-request-id': 'req_01kesv4x14eptanxe44mfxndqd',
   via: '1.1 google',
   'cf-cache-status': 'DYNAMIC',
   'set-cookie': '__cf_bm=AwxBo2U51m8khHAx3TPmbXRHVrDRBk3G7nLSZBsEFKA-1768246375-1.0.1.1-2ipUeEYn.NixbCYYMofApTbyS02oKlDZX8WpGN3Xl8twVthSMx8lHd1GBZS0bQamwtWOG6yiverHhfzujUw0EQteGSjMzKSC9K
iQ54ioBxE; path=/; expires=Mon, 12-Jan-26 20:02:55 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
   'strict-transport-security': 'max-age=15552000',
   server: 'cloudflare',
   'alt-svc': 'h3=":443"; ma=86400'
 },
 requestID: 'req_01kesv4x14eptanxe44mfxndqd',
 error: [Object],
 code: undefined,
 param: 'response_format',
 type: 'invalid_request_error'
}
Error: tool call validation failed: parameters for tool web_search did not match schema: errors: [missing properties: 'type']
   at bf.f [as iterator] (.next/server/chunks/607.js:20:27895)
   at async M.streamText (.next/server/chunks/136.js:1:3089)
   at async i.research (.next/server/chunks/641.js:541:227)
   at async m.searchAsync (.next/server/app/api/chat/route.js:1:13218) {
 status: undefined,
 headers: Headers {
   date: 'Mon, 12 Jan 2026 19:34:06 GMT',
   'content-type': 'text/event-stream',
   'transfer-encoding': 'chunked',
   connection: 'keep-alive',
   'cf-ray': '9bcf16e19c5f4292-VCP',
   'cache-control': 'no-cache',
   vary: 'Origin',
   'x-groq-region': 'bra',
   'x-ratelimit-limit-requests': '500000',
   'x-ratelimit-limit-tokens': '250000',
   'x-ratelimit-remaining-requests': '499999',
   'x-ratelimit-remaining-tokens': '248367',
   'x-ratelimit-reset-requests': '172ms',
   'x-ratelimit-reset-tokens': '391ms',
   'x-request-id': 'req_01kesv7291eqf85rpfh3w5xf4g',
   via: '1.1 google',
   'cf-cache-status': 'DYNAMIC',
   'set-cookie': '__cf_bm=.Q_selQ6IHgefUKPO2khiqMIVmt4hgIfKU0_hELNXwg-1768246446-1.0.1.1-AR4vvKo1mH_3oyemG_ZCevMdq8DuObZOaZ5E5BvAJ.z_b7svpibeUr2eU1ECwYdhn9c9EnOUHqf0Ii6eJ5FDaHb8GpuBYLQJtw
tpIswdQnk; path=/; expires=Mon, 12-Jan-26 20:04:06 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
   'strict-transport-security': 'max-age=15552000',
   server: 'cloudflare',
   'alt-svc': 'h3=":443"; ma=86400'
 },
 requestID: 'req_01kesv7291eqf85rpfh3w5xf4g',
 error: [Object],
 code: 'tool_use_failed',
 param: undefined,
 type: 'invalid_request_error'
}
⨯ unhandledRejection:  Error: tool call validation failed: parameters for tool web_search did not match schema: errors: [missing properties: 'type']
   at bf.f [as iterator] (.next/server/chunks/607.js:20:27895)
   at async M.streamText (.next/server/chunks/136.js:1:3089)
   at async i.research (.next/server/chunks/641.js:541:227)
   at async m.searchAsync (.next/server/app/api/chat/route.js:1:13218) {
 status: undefined,
 headers: Headers {
   date: 'Mon, 12 Jan 2026 19:34:06 GMT',
   'content-type': 'text/event-stream',
   'transfer-encoding': 'chunked',
   connection: 'keep-alive',
   'cf-ray': '9bcf16e19c5f4292-VCP',
   'cache-control': 'no-cache',
   vary: 'Origin',
   'x-groq-region': 'bra',
   'x-ratelimit-limit-requests': '500000',
   'x-ratelimit-limit-tokens': '250000',
   'x-ratelimit-remaining-requests': '499999',
   'x-ratelimit-remaining-tokens': '248367',
   'x-ratelimit-reset-requests': '172ms',
   'x-ratelimit-reset-tokens': '391ms',
   'x-request-id': 'req_01kesv7291eqf85rpfh3w5xf4g',
   via: '1.1 google',
   'cf-cache-status': 'DYNAMIC',
   'set-cookie': '__cf_bm=.Q_selQ6IHgefUKPO2khiqMIVmt4hgIfKU0_hELNXwg-1768246446-1.0.1.1-AR4vvKo1mH_3oyemG_ZCevMdq8DuObZOaZ5E5BvAJ.z_b7svpibeUr2eU1ECwYdhn9c9EnOUHqf0Ii6eJ5FDaHb8GpuBYLQJtw
tpIswdQnk; path=/; expires=Mon, 12-Jan-26 20:04:06 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
   'strict-transport-security': 'max-age=15552000',
   server: 'cloudflare',
   'alt-svc': 'h3=":443"; ma=86400'
 },
 requestID: 'req_01kesv7291eqf85rpfh3w5xf4g',
 error: [Object],
 code: 'tool_use_failed',
 param: undefined,
 type: 'invalid_request_error'
}

[pfjarschel]

I won't immediately mark this as closed, because there might still be something going on, or might be helpful for someone else.
But I saw in another thread that someone was having some issues, related with json stuff, and @ItzCrazyKns suggested a fresh install. Well, that worked for me! So I will wait for any comments here for a bit, and then I can mark this issue as closed.

Just some more info on the fresh install procedure: I did not have to prune containers, volumes, images, or anything. I simply created another folder, copied the compose.yaml file and started it up.

[pfjarschel]

Ok, new information. Using the Kimi K2 Instruct model via Groq still causes the tool calling error:

Error: tool call validation failed: parameters for tool web_search did not match schema: errors: [missing properties: 'type']
    at bf.f [as iterator] (.next/server/chunks/607.js:20:27895)
    at async M.streamText (.next/server/chunks/136.js:1:3089)
    at async i.research (.next/server/chunks/641.js:541:227)
    at async m.searchAsync (.next/server/app/api/chat/route.js:1:13218) {
  status: undefined,
  headers: Headers {
    date: 'Tue, 13 Jan 2026 14:36:12 GMT',
    'content-type': 'text/event-stream',
    'transfer-encoding': 'chunked',
    connection: 'keep-alive',
    'cf-ray': '9bd59fd64f824287-VCP',
    'cache-control': 'no-cache',
    vary: 'Origin',
    'x-groq-region': 'bra',
    'x-ratelimit-limit-requests': '500000',
    'x-ratelimit-limit-tokens': '250000',
    'x-ratelimit-remaining-requests': '499999',
    'x-ratelimit-remaining-tokens': '248363',
    'x-ratelimit-reset-requests': '172ms',
    'x-ratelimit-reset-tokens': '392ms',
    'x-request-id': 'req_01kevwj8fzfdxsjnm4b99re2q8',
    via: '1.1 google',
    'cf-cache-status': 'DYNAMIC',
    'set-cookie': '__cf_bm=6L3XezKqGxiZbB2kB2zhKq7oIxIkvJ5kpTbt0MX7nkg-1768314972-1.0.1.1-Kx1vp6OKWlxC_ROCmOJiawvN9sN8paz2ykuGN1JKXuCaJGH0K3TJCFuEy4Oet441xdEh6bILMEykjeqn1W8ajkFEDuUCjdnx1uZa93YnmUs; path=/; expires=Tue, 13-Jan-26 15:06:12 GMT; domain=.groq.com; HttpOnly; Secure; SameSite=None',
    'strict-transport-security': 'max-age=15552000',
    server: 'cloudflare',
    'alt-svc': 'h3=":443"; ma=86400'
  },
  requestID: 'req_01kevwj8fzfdxsjnm4b99re2q8',
  error: [Object],
  code: 'tool_use_failed',
  param: undefined,
  type: 'invalid_request_error'
}

So this issue is still there.

[andreyshirokoff]

I've got the such trouble on Anthropic models. Confirm.

[blueemp]

问题分析

我已经分析了代码，找到了两个相关的问题：

问题 1：json_schema 响应格式错误

根本原因：
在 src/lib/models/providers/openai/openaiLLM.ts 的 generateObject 方法中，使用了 zodResponseFormat 来生成结构化输出：

response_format: zodResponseFormat(input.schema, "object"),

这个 API 调用会向 OpenAI API 发送 response_format: { type: "json_schema", ... }。

问题：
Groq API 可能不完全支持 json_schema 格式，或者对某些模型的支持有限制。即使 Groq 文档中列出了支持 json_schema 的模型，实际 API 行为可能不同。

问题 2：web_search 工具 schema 缺少 type 属性

根本原因：
在 src/lib/models/providers/openai/openaiLLM.ts 的 generateText 和 streamText 方法中，工具参数生成使用了：

openaiTools.push({
  type: "function",
  function: {
    name: tool.name,
    description: tool.description,
    parameters: z.toJSONSchema(tool.schema),
  },
});

问题：
z.toJSONSchema() 生成的 schema 可能不包含 Groq 要求的 type 属性，导致工具验证失败。

修复建议

修复 1：为 Groq 添加回退机制

在 src/lib/models/providers/groq/groqLLM.ts 中覆盖 generateObject 方法，使用 JSON 模式而不是 json_schema：

import OpenAILLM from "../openai/openaiLLM";
import { zodToJSONSchema } from "zod-to-json-schema";

class GroqLLM extends OpenAILLM {
  async generateObject<T>(input: GenerateObjectInput): Promise<T> {
    // 使用 JSON 模式而不是 json_schema
    const response = await this.openAIClient.chat.completions.create({
      messages: this.convertToOpenAIMessages(input.messages),
      model: this.config.model,
      temperature: input.options?.temperature ?? this.config.options?.temperature ?? 1.0,
      top_p: input.options?.topP ?? this.config.options?.topP,
      max_completion_tokens: input.options?.maxTokens ?? this.config.options?.maxTokens,
      response_format: { type: "json_object" },
    });

    if (response.choices && response.choices.length > 0) {
      try {
        const content = response.choices[0].message.content;
        return input.schema.parse(JSON.parse(content!)) as T;
      } catch (err) {
        throw new Error(`Error parsing response from Groq: ${err}`);
      }
    }

    throw new Error("No response from Groq");
  }
}

export default GroqLLM;

修复 2：确保工具 schema 包含 type 属性

更新 src/lib/models/providers/openai/openaiLLM.ts 中的工具参数生成：

input.tools?.forEach((tool) => {
  const schema = z.toJSONSchema(tool.schema);
  
  // 确保 schema 有 type 属性
  if (!schema.type) {
    schema.type = "object";
  }
  
  openaiTools.push({
    type: "function",
    function: {
      name: tool.name,
      description: tool.description,
      parameters: schema,
    },
  });
});

修复 3：添加错误处理和重试逻辑

在 Groq 提供者中添加模型能力检测和回退逻辑。

临时解决方案

1. 降级到 1.11.2：如报告所述，降级可以暂时解决问题
2. 使用其他提供商：切换到 OpenAI 或其他完全兼容的提供商
3. 禁用结构化输出：修改代码以使用纯 JSON 解析而不是 json_schema

需要的信息

为了提供更精确的修复，需要：

1. Groq API 的具体错误响应
2. 使用的确切模型 ID
3. Groq API 版本信息

建议联系 Groq 支持确认：

• 哪些模型支持 json_schema
• json_schema 的正确使用方式
• 工具 schema 的具体要求

[sjiampojamarn]

Facing the same issues. Confirmed.