thanks !

C

Celestial RoseOP•11/26/25, 5:18 PM

sent

S

Sb•11/27/25, 3:38 AM

Can I use different models than the current provided.
Like wan 2.2 or veo 3 or other open source model by paying for the compute or neurons?

S

Sb•11/27/25, 3:41 AM

Cloudflare is already using GPU for other models like flux and llama,
Will they also provide us with the direct usage?

M

muskan•11/27/25, 9:44 AM

Hi! I wanted to confirm something about Cloudflare AI. Does it support parsing PDF files directly?
Right now, it looks like I must extract the text from the PDF myself before sending it to the AI model. Is there any built-in support (or upcoming support) for uploading a PDF and having the AI read/parse it automatically?

A

Adnan A.•11/27/25, 1:23 PM

Hi,

I’m encountering an issue with

Ai-Search (Auto RAG)

Ai-Search (Auto RAG)

Ai-Search (Auto RAG)

Ai-Search (Auto RAG) with folder-based filtering using the

gt

gt

gt

gt,

gte

gte

gte

gte, and

eq

eq

eq

eq filter types. Filtering works as expected when targeting a top-level folder, but fails when specifying a sub-folder.

Below is an example payload that works and returns all files inside

customer-a

customer-a

customer-a

customer-a as well as its sub-folders:

{
  "query": "Cat",
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a//"
  }
}

{
  "query": "Cat",
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a//"
  }
}

{
  "query": "Cat",
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a//"
  }
}

{
  "query": "Cat",
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a//"
  }
}

However, the following payloads do not return any results, although the specified sub-folder exists:

{
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gt",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "eq",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "eq",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "eq",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "eq",
    "key": "folder",
    "value": "customer-a/sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder//"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder//"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder//"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a/sub-folder//"
  }
}

The following variant does work, but only returns items from the folder itself — not from its deeper sub-folders:

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a//sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a//sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a//sub-folder"
  }
}

{
  "filters": {
    "type": "gte",
    "key": "folder",
    "value": "customer-a//sub-folder"
  }
}

What I need

I would like to search inside a specific sub-folder, ideally including its nested sub-folders, similar to how filtering works for the top-level folder.

Could you please advise on the correct filtering method or confirm whether this is a known limitation or bug?

Thank you in advance for your assistance.

Mmuskan Hi! I wanted to confirm something about Cloudflare AI. Does it support parsing P...

S

Sb•11/27/25, 2:05 PM

Pdf to markdown in cf

Mmuskan Hi! I wanted to confirm something about Cloudflare AI. Does it support parsing P...

B

blah•11/27/25, 3:13 PM

google has stuff like that out of the box that you can work into your pipeline.

M

muskan•11/28/25, 6:01 AM

Hi! Thanks for the replies
I have one follow-up question:

I want to upload the whole PDF file directly to the AI model and get back a clean, readable JSON — without having to parse the PDF myself first.

Basically:
No pre-processing
No pdf-parse / PDF.js / external extraction
Just: Upload PDF → AI extracts → returns structured output

Is there a recommended Cloudflare AI model / workflow that already supports this type of direct PDF→JSON parsing?
A model similar to “document AI” (like Google’s), where you upload a PDF and get structured data back?

Thanks again!

Mmuskan Hi! Thanks for the replies I have one follow-up question: 👉 I want to upload t...

S

Sb•11/28/25, 12:49 PM

Check cloudflare pdf to markdown for worker ai.
It does exactly that.

R

raxrb•11/28/25, 9:06 PM

Does the wokers AI has cold start issue?

X

xtudos•11/28/25, 10:15 PM

hello, i'm trying to look for

@cf/meta/llama-3.1-8b-instruct-fast

@cf/meta/llama-3.1-8b-instruct-fast

@cf/meta/llama-3.1-8b-instruct-fast

@cf/meta/llama-3.1-8b-instruct-fast but can't find it on the dashboard

it exists on the docs, but not on the dashboard nor API. I was trying to look for princing info

https://developers.cloudflare.com/workers-ai/models/llama-3.1-8b-instruct-fast/

T

tkpop777•11/29/25, 9:52 PM

I am trying to use Flux Schnell text to image model from my worker and I am encountering censorship, not only that but I get censorship errors even when I didn't enter anything spicy in the prompts like just the word "Elephant" triggers:

"AiError: Input prompt contains NSFW content."

and the other thing is, what does this model accept 2 params (prompt and steps)? its unusable since I can't even ask for a specific size.

Any way to disable the NSFW filter and accept more params?

V

Vlad Lykhonis•12/1/25, 3:31 PM

Hi all, did anyone try to stream for openai/gpt-oss models? I keep getting Neteork connection lost. But only for stream. I tried with llama-3.3.-70b and stream works fine.

VVlad Lykhonis Hi all, did anyone try to stream for openai/gpt-oss models? I keep getting Neteo...

A

ajgeiss0702•12/1/25, 5:08 PM

streaming works fine for me. gpt-oss has a completly different output schema from the other models, maybe thats part of the issue?

Aajgeiss0702 streaming works fine for me. gpt-oss has a completly different output schema fro...

V

Vlad Lykhonis•12/1/25, 7:40 PM

Are you using workers typescript binding? I do that, and readable stream output gets me only one 41 byte chunk stating network connection lost. This happens when I apply stream flag.

Are you doing the same and you get streaming chunks fine?

A

ajgeiss0702•12/1/25, 7:40 PM

im not, im using the http api

V

Vlad Lykhonis•12/1/25, 7:41 PM

Right, Iâm referring to workers ai binding. Not http. And output schema shouldnât matter since itâs streaming such as itâs readable stream response

A

acusti•12/1/25, 11:05 PM

@Vlad Lykhonis i havenât been able to get streaming to work with the gpt-oss models and the workers AI binding. the last mention i found about it in this channel was from August 27 from @samjs, where he mentioned that theyâre working on it and hope to have an update soon: workers-ai

Aacusti @Vlad Lykhonis i havenât been able to get streaming to work with the gpt-oss m...

V

Vlad Lykhonis•12/2/25, 5:58 AM

Thank you.

V

Vlad Lykhonis•12/2/25, 6:00 AM

On another topic regarding usage storage and logging.

My understanding is that workers ai do NOT store any data or logs for inputs and outputs of any of the models involved.

Of course, I’m talking about direct workers AI no gateways etc.

I’m referring to this section of https://developers.cloudflare.com/workers-ai/platform/data-usage/

Your Customer Content for Workers AI may be stored by Cloudflare if you specifically use a storage service (e.g., R2, KV, DO, Vectorize, etc.) in conjunction with Workers AI.

Would that be correct?

Cloudflare Docs

Data usage

Cloudflare processes certain customer data in order to provide the Workers AI service, subject to our Privacy Policy and Self-Serve Subscription Agreement or Enterprise Subscription Agreement (as applicable).

S

Sb•12/2/25, 8:05 AM

Hey cloudflare

I wanted to ask you that after the replicate partnership,
Will we be able to use the cloudflare neurons to run replicate models ?

Original message was deleted

D

DaniFoldi•12/2/25, 9:09 AM

Yep, ty

C

Celestial RoseOP•12/2/25, 5:17 PM

hey channel ! any idea if or when a fine-tuning pipeline product will see the light ? like ai search but for fine tuning, that would be so dope

C

Celestial RoseOP•12/2/25, 5:25 PM

is the LoRA models up to date on Workers AI ? Llama4 isnt in the list but im pretty sure it can use LoRA weights

A

Anton•12/2/25, 6:05 PM

I'm getting "error code: 1031" repeatedly trying to convert a jpg to markdown using

ai.toMarkdown()

ai.toMarkdown()

ai.toMarkdown()

ai.toMarkdown() (docs), that error code doesn't seem to be documented. Any idea why it's failing?

AAnton I'm getting "error code: 1031" repeatedly trying to convert a jpg to markdown us...

A

acusti•12/2/25, 10:33 PM

that happens to me multiple times a day using wrangler in local dev and i have to use

wrangler login

wrangler login

wrangler login

wrangler login via my package manager (e.g.

npx wrangler login

npx wrangler login

npx wrangler login

npx wrangler login or

bun wrangler login

bun wrangler login

bun wrangler login

bun wrangler login in my case) to get it to work again. it’s always error code: 1031 that tells me i need to do that.

to be clear, i don’t think that’s intended to work that way, so there’s probably something wrong with my setup that causes that behavior.

C

captain•12/3/25, 10:23 PM

been reviewing the new flux 2, and comparing using on workers vs fal.ai, i would rather just use it in workers to make things less complex. However it seems like using it on fal.ai if i campare the demo promt in the workers docs: 16 steps + 1024x1024. That demo would be half the price on fal, does anybody know if i am calculating this incorrectly or if i am not properly understanding the value of the worker request. I am typically used to cloudflare costs being lower but in this case it doesn't seem so

I

Ichimiya•12/5/25, 1:34 AM

Is wokers AI considering launching the z-image-turbo image generation model?

R

Reux•12/8/25, 6:50 PM

I've been using Flux 2 Dev, and there honestly seems to be a serious issue with the filter for prompts and input images, there are way too many false positives, It's flagging completely legitimate images and prompts that are the exact opposite of inappropriate. The false positives are so frequent that it's impossible to use the model in a real-world application, please take a look at the filter

R

rjrobben•12/9/25, 2:34 AM

is there any plan to add the model

bge-reranker-v2-m3

bge-reranker-v2-m3

bge-reranker-v2-m3

bge-reranker-v2-m3 ? would be useful to plenty of folks

B

Bernd•12/9/25, 1:00 PM

What is the correct page to see the pricings?

Take @cf/google/embeddinggemma-300m:

The model page says nothing: https://developers.cloudflare.com/workers-ai/models/embeddinggemma-300m/
The workers ai platform page also says nothing: https://developers.cloudflare.com/workers-ai/platform/pricing/
But then the models overview that doesn't have descriptions, shown a neuron price, but unlike other models no price per million in USD: https://dash.cloudflare.com/123456/ai/workers-ai/models

Does this mean the model can only be used via neurons? Or is this a documenation issue?

Cloudflare Docs

embeddinggemma-300m

EmbeddingGemma is a 300M parameter, state-of-the-art for its size, open embedding model from Google, built from Gemma 3 (with T5Gemma initialization) and the same research and technology used to create Gemini models. EmbeddingGemma produces vector representations of text, making it well-suited for search and retrieval tasks, including classifica...

Cloudflare Docs

Pricing

Workers AI is included in both the Free and Paid Workers plans and is priced at $0.011 per 1,000 Neurons.

R

Robin•12/9/25, 2:49 PM

Hey folks

,

has anyone managed to get streaming working with gpt-oss models on workers?
if so please could you share your setup, been banging my head against the wall trying to get this to work.

D

Derfog•12/11/25, 2:32 PM

Good evening everyone, I would like recommendations of better ais from worker ai, to process data, etc, I wanted to start on Mistral models but I would like your opinions, thank you

B

blindChicken•12/12/25, 3:49 PM

I would just like to make some noise, in the hopes of encouraging the fine folks at CF, to provide us with an OCR model such as DeepSeek-OCR.

V

Vikash•12/13/25, 4:45 AM

Workers AI not supported on Vercel AI sdk providers?

const aiGateway = createAiGateway({
    accountId: env.CF_ACCOUNT_ID,
    gateway: GATEWAY_ID,
    apiKey: env.CF_API_KEY,
    options: {
      cacheTtl: 3600,
    },
  });

  const workersai = createWorkersAI({ binding: env.AI });
  const model = workersai("@cf/meta/llama-3.1-8b-instruct", {
    safePrompt: true,
  });

  const { text } = await generateText({
    model: aiGateway(model),
    prompt: "Write a multi-part greeting.",
  });

const aiGateway = createAiGateway({
    accountId: env.CF_ACCOUNT_ID,
    gateway: GATEWAY_ID,
    apiKey: env.CF_API_KEY,
    options: {
      cacheTtl: 3600,
    },
  });

  const workersai = createWorkersAI({ binding: env.AI });
  const model = workersai("@cf/meta/llama-3.1-8b-instruct", {
    safePrompt: true,
  });

  const { text } = await generateText({
    model: aiGateway(model),
    prompt: "Write a multi-part greeting.",
  });

const aiGateway = createAiGateway({
    accountId: env.CF_ACCOUNT_ID,
    gateway: GATEWAY_ID,
    apiKey: env.CF_API_KEY,
    options: {
      cacheTtl: 3600,
    },
  });

  const workersai = createWorkersAI({ binding: env.AI });
  const model = workersai("@cf/meta/llama-3.1-8b-instruct", {
    safePrompt: true,
  });

  const { text } = await generateText({
    model: aiGateway(model),
    prompt: "Write a multi-part greeting.",
  });

const aiGateway = createAiGateway({
    accountId: env.CF_ACCOUNT_ID,
    gateway: GATEWAY_ID,
    apiKey: env.CF_API_KEY,
    options: {
      cacheTtl: 3600,
    },
  });

  const workersai = createWorkersAI({ binding: env.AI });
  const model = workersai("@cf/meta/llama-3.1-8b-instruct", {
    safePrompt: true,
  });

  const { text } = await generateText({
    model: aiGateway(model),
    prompt: "Write a multi-part greeting.",
  });

VVikash Workers AI not supported on Vercel AI sdk providers? ``` const aiGateway = crea...

P

Palash•12/15/25, 4:13 PM

workers-ai-provider natively supports using ai gateway, you should be able to pass

createWorkersAI({ binding: env.AI, gateway: { id: GATEWAY_ID }});

createWorkersAI({ binding: env.AI, gateway: { id: GATEWAY_ID }});

createWorkersAI({ binding: env.AI, gateway: { id: GATEWAY_ID }});

createWorkersAI({ binding: env.AI, gateway: { id: GATEWAY_ID }});

M

Mukesh Patidar•12/17/25, 8:00 AM

not able to see preview, getting this issue
can you please help me with that.

MMukesh Patidar not able to see preview, getting this issue can you please help me with that.

I

Isaac McFadyen•12/17/25, 1:53 PM

?crossposting

S

SuperHelpflare•12/17/25, 1:53 PM

Please do not post your question in multiple channels/post it multiple times per the rules at No Access. It creates confusion for people trying to help you and doesn't get your issue or question solved any faster.

니

니꼬•12/18/25, 11:31 PM

Anyone knows how I can access 'env.X' from from

server.tool

server.tool

server.tool

server.tool callback?

import { createMcpHandler } from "agents/mcp";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";

const server = new McpServer({
  name: "Hello MCP Server",
  version: "1.0.0",
});

server.tool(
  "hello",
  "Returns a greeting message",
  { name: z.string().optional() },
  async ({ name }) => {
    return {
      content: [
        {
          text: `Hello, ${name ?? "World"}!`,
          type: "text",
        },
      ],
    };
  },
);

export default {
  fetch: async (request: Request, env: Env, ctx: ExecutionContext) => {
    const handler = createMcpHandler(server);
    return handler(request, env, ctx);
  },
};

import { createMcpHandler } from "agents/mcp";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";

const server = new McpServer({
  name: "Hello MCP Server",
  version: "1.0.0",
});

server.tool(
  "hello",
  "Returns a greeting message",
  { name: z.string().optional() },
  async ({ name }) => {
    return {
      content: [
        {
          text: `Hello, ${name ?? "World"}!`,
          type: "text",
        },
      ],
    };
  },
);

export default {
  fetch: async (request: Request, env: Env, ctx: ExecutionContext) => {
    const handler = createMcpHandler(server);
    return handler(request, env, ctx);
  },
};

import { createMcpHandler } from "agents/mcp";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";

const server = new McpServer({
  name: "Hello MCP Server",
  version: "1.0.0",
});

server.tool(
  "hello",
  "Returns a greeting message",
  { name: z.string().optional() },
  async ({ name }) => {
    return {
      content: [
        {
          text: `Hello, ${name ?? "World"}!`,
          type: "text",
        },
      ],
    };
  },
);

export default {
  fetch: async (request: Request, env: Env, ctx: ExecutionContext) => {
    const handler = createMcpHandler(server);
    return handler(request, env, ctx);
  },
};

import { createMcpHandler } from "agents/mcp";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";

const server = new McpServer({
  name: "Hello MCP Server",
  version: "1.0.0",
});

server.tool(
  "hello",
  "Returns a greeting message",
  { name: z.string().optional() },
  async ({ name }) => {
    return {
      content: [
        {
          text: `Hello, ${name ?? "World"}!`,
          type: "text",
        },
      ],
    };
  },
);

export default {
  fetch: async (request: Request, env: Env, ctx: ExecutionContext) => {
    const handler = createMcpHandler(server);
    return handler(request, env, ctx);
  },
};

니니꼬 Anyone knows how I can access 'env.X' from from `server.tool` callback? ```html ...

I

Isaac McFadyen•12/19/25, 2:15 AM

Try using the new(ish) global env import: https://developers.cloudflare.com/workers/runtime-apis/bindings/#importing-env-as-a-global

R

Reux•12/20/25, 1:38 PM

There is an issue with flux 2 dev, when you pass an input image and you specify a different width and height, the specified size is ignored and keeps the input image size

S

scotto•12/20/25, 10:36 PM

was comparing gemma embedding agains other providers as we need other providers for fallback, but embeddings are different

S

scotto•12/20/25, 10:36 PM

Summary: Embedding Comparison Results

| Comparison | Text 1 (English) | Text 2 (Spanish product) |
|---------------------------|------------------|--------------------------|
| HuggingFace vs DeepInfra | 0.9999 ✓ | 0.9999 ✓ |
| HuggingFace vs Cloudflare | 0.7267 | 0.5969 |
| Cloudflare vs DeepInfra | 0.7252 | 0.5967 |

G

Ganesh_222•12/22/25, 6:34 AM

Hello team, Query rewrite not working, can you please check and confirm?

C

Celestial RoseOP•12/22/25, 10:59 AM

loving the new UI for AI Search so far !!!

C

Celestial RoseOP•12/22/25, 10:59 AM

kudos !

N

ngghghfjgnghgngnhhghngn•12/22/25, 11:42 AM

Hello team, Query rewrite not working,

S

Sergey Dudko•12/22/25, 2:32 PM

Hi guys. From time to time, I get empty content from the

toMarkdown

toMarkdown

toMarkdown

toMarkdown method (rest api). The request looks like it was successful: I get document metadata, Contents, and numbered pages in response, but the pages themselves contain no data. However, there is data on the pdf's pages (text and images). Do you have any explanations/solutions for this?

# 2.pdf
## Metadata
- PDFFormatVersion=1.4
- IsLinearized=false
- IsAcroFormPresent=false
...

## Contents
### Page 1




### Page 2




### Page 3


...

# 2.pdf
## Metadata
- PDFFormatVersion=1.4
- IsLinearized=false
- IsAcroFormPresent=false
...

## Contents
### Page 1




### Page 2




### Page 3


...

# 2.pdf
## Metadata
- PDFFormatVersion=1.4
- IsLinearized=false
- IsAcroFormPresent=false
...

## Contents
### Page 1




### Page 2




### Page 3


...

# 2.pdf
## Metadata
- PDFFormatVersion=1.4
- IsLinearized=false
- IsAcroFormPresent=false
...

## Contents
### Page 1




### Page 2




### Page 3


...

IIsaac McFadyen Try using the new(ish) global env import: https://developers.cloudflare.com/work...

니

니꼬•12/22/25, 2:48 PM

thanks!

What I need

Similar Threads

What I need