Cloudflare Developers•8mo ago

check readme

✎✎ℳ㎕Joy485@新歲錦時 check readme

Isaac McFadyen•5/17/25, 4:14 PM

Yes but this is the Workers AI channel, not for general Workers apps.

✎

✎ℳ㎕Joy485@新歲錦時OP•5/17/25, 4:16 PM

why tf would we be the traditional way

✎

✎ℳ㎕Joy485@新歲錦時OP•5/17/25, 4:18 PM

with Poor adaptability and a deployment cycle as slow as a turtle

Isaac McFadyen•5/17/25, 4:19 PM

I'm not sure what you mean but what I was trying to say is that you posted an app unrelated to Cloudflare Workers AI in a channel meant for only Cloudflare Workers AI questions. Please avoid doing that in the future.

yinxingmaiming6409•5/20/25, 3:29 AM

Why does AI have neuron prices and token prices? What price should I consider for my worker paid? Or should we charge double the price?

yinxingmaiming6409•5/20/25, 3:32 AM

@How to use Woekr AI to extract the vector of r2 image from CF/Microsoft/Resnet-50 and save it to vectordb? Do you have any examples?

yinxingmaiming6409•5/20/25, 3:39 AM

Yyinxingmaiming6409 Why does AI have neuron prices and token prices? What price should I consider fo...

Viktor•5/20/25, 12:38 PM

They are the same. Different models use a different amount of neurons per token, so it tells you both prices - the neuron price (which is the same for all models) and the price per token per model (which differs, as models have different neurons/tokens). Cloudflare used to bill only per neuron, but as every ai provider has price per token, they added that, too

VViktor They are the same. Different models use a different amount of neurons per token,...

yinxingmaiming6409•5/20/25, 3:02 PM

Resnet50 can extract embeddings, can't you?

Yyinxingmaiming6409 Resnet50 can extract embeddings, can't you?

Viktor•5/20/25, 3:09 PM

It doesn't look like that. I'd say it does return a list of things it recognised, and not an embedding

MSL•5/20/25, 3:36 PM

Hey, has there been any considerations for having the Workers option be limited to storage in the EU to abide by GDPR-Compliance? Something like the " Cloudflare Data Localization Suite" but without having to be inside an Enterprise suite

MMSL Hey, has there been any considerations for having the Workers option be limited ...

Viktor•5/21/25, 9:08 AM

Cloudflare is certified by eu us data privacy framework, it is GDPR compliant. But for some storage options (like R2, D1) you can already set a region. You posted this question in workers-ai and asked about storage - do you mean specifically vectorize?

VViktor Cloudflare is certified by eu us data privacy framework, it is GDPR compliant. B...

MSL•5/21/25, 9:36 AM

No, I am specifically talking about Workers, since they run on all network locations that can potentially be a GDPR issue for data handling,
It’s not as easy as them being certified, if a worker parses sensitive data but is hosted outside of the EU that can be an issue even with the Cloudflare DPA should cover it, it depends on the data you’re handling

MSL•5/21/25, 9:43 AM

Searching for it also shows people arguing and general confusion regarding the DPA with no official reply from Cloudflare except “You can sign a DPA with us” but Pages and Workers are specifically written to not be an option to have Data Localization only in the EU. Which is where I doubt the management of GDPR is sufficient, as the processor I would still be the one with fines if some data is processed outside of the EU

Mochni•5/21/25, 11:54 PM

Mochni•5/21/25, 11:55 PM

It says I used 691.2/10k neurons, but the chart says I used 16.93k neurons

Mochni•5/21/25, 11:55 PM

What am I missing here?

Object•5/23/25, 9:24 PM

Hello, reporting an issue with Gemma 3 in Workers AI, or maybe just the documentation about it.

The Gemma 3 model is not behaving as documented. Two issues so far:

(minor but a bit annoying) "@cf/google/gemma-3-12b-it" is not included in
```
wrangler types
```
```
wrangler types
```
the following code chokes, even though correct per the official doc page (https://developers.cloudflare.com/workers-ai/models/gemma-3-12b-it/) . The part it dislikes is
```
messages[1].content
```
```
messages[1].content
```
being an object instead of string. The string format only works for text, however, not image, i.e.
```
{role: 'user', content: dataUri}
```
```
{role: 'user', content: dataUri}
```
will not choke but Gemma will completely hallucinate the actual content of the image, while recognizing the same image just fine if uploaded in aistudio.google.com.

const res = await fetch(

https://cataas.com/cat

https://cataas.com/cat

);
const blob = await res.arrayBuffer();
const imageData = [...new Uint8Array(blob)];

const base64 = btoa(String.fromCharCode(...imageData));
const dataUri =

data:image/jpeg;base64,${base64}

data:image/jpeg;base64,${base64}

;

const response = await env.AI.run(
"@cf/google/gemma-3-12b-it",
{
messages: [
{role: 'system', content: 'describe uploaded images'},
{role: 'user',
content: {
type: 'image', // tried with and without this line
image_url: {url: dataUri}
}
}
]
}).catch(e => {
console.error(

AI.run error

AI.run error

, e);
throw e;
});

return new Response(JSON.stringify(response), {
headers: [['content-type', 'application/json']]
});

Object•5/23/25, 9:28 PM

Here is the error the run API returns:
✘ [ERROR] AI error InferenceUpstreamError [AiError]: 3030: 3 validation errors for ValidatorIterator

0.typed-dict
Input should be a valid dictionary [type=dict_type, input_value='image_url', input_type=str]
For further information visit https://errors.pydantic.dev/2.10/v/dict_type
0.typed-dict
Input should be a valid dictionary [type=dict_type, input_value='image_url', input_type=str]
For further information visit https://errors.pydantic.dev/2.10/v/dict_type
0.typed-dict
Input should be a valid dictionary [type=dict_type, input_value='image_url', input_type=str]
For further information visit https://errors.pydantic.dev/2.10/v/dict_type
at Ai._parseError (cloudflare-internal:ai-api:102:24)
at async Ai.run (cloudflare-internal:ai-api:82:19)
at response (file:///<...>/classify.ts:58:20)
at async Object.fetch
(file:///<...>node_modules/src/src/Router.ts:58:20)
at async jsonError
(file:///<...>/node_modules/wrangler/templates/middleware/middleware-miniflare3-json-error.ts:22:10)
at async drainBody
(file:///<...>node_modules/wrangler/templates/middleware/middleware-ensure-req-body-drained.ts:5:10)

SODaLC•5/25/25, 2:59 PM

Im current running into an error about "JSON Model couldn't be met" . Is there a way to tell why this is happening or what it is having trouble with with some sort of verbose mode? At the moment, it very inconsistent.

zensin•5/27/25, 4:07 AM

Please add Devstral-Small-2505 if possible
https://huggingface.co/mistralai/Devstral-Small-2505
https://huggingface.co/unsloth/Devstral-Small-2505

Dev•5/27/25, 5:02 AM

Hi, why can't we use BGE-M3 embedding model (@cf/baai/bge-m3) to create vector index?

Only these model are available to choose from: [ "@cf/baai/bge-small-en-v1.5", "@cf/baai/bge-base-en-v1.5", "@cf/baai/bge-large-en-v1.5", "openai/text-embedding-ada-002", "cohere/embed-multilingual-v2.0" ]

eli-front•5/27/25, 11:16 PM

Has anyone successfully setup tool call streaming with any of the function calling models? I can get tool calling to work when it's not streaming.

DDev Hi, why can't we use BGE-M3 embedding model (@cf/baai/bge-m3) to create vector i...

Viktor•5/28/25, 10:46 AM

I have no idea what you mean, but you can create a vector index with length = 1024 via wrangler, and store vectors created via bge-m3 in it

Eeli-front Has anyone successfully setup tool call streaming with any of the function calli...

Viktor•5/28/25, 10:49 AM

I buffered the first bytes and checked if they contain json or xml (depending on the model). If it does not start with

{"

{"

<function

<function

(it was I think) it's not a function call, else it is one. In theory @cloudflare/ai-utils can do that, but it calls the model twice (once without streaming and in case it does not call a tool - again with streaming), so I didn't want to use it

GitHub

GitHub - cloudflare/ai-utils: Developer toolkit that makes it simpl...

Developer toolkit that makes it simple to build with the Workers AI platform. - cloudflare/ai-utils

plgingras88•5/29/25, 2:09 AM

Using toMarkdown on a jpeg in my R2 bucket right now returns AiError: 3043: Internal server error. I changed nothing and was working fine on Sunday, any idea ?

Pplgingras88 Using toMarkdown on a jpeg in my R2 bucket right now returns AiError: 3043: Inte...

david-thyresson•5/29/25, 11:47 AM

Have also seen this since I tried using it yesterday - pdf and csv and xls all converted fine. Just the images error for me. Am using current 4.17.0 of wrangler. Am running local dev.

max•5/29/25, 6:50 PM

Hi! Does anyone have any idea why I am getting this error from hitting the Whisper AI endpoint?

{"errors":[{"message":"Error: Network connection lost.","code":6001}],"success":false,"result":{},"messages":[]}

{"errors":[{"message":"Error: Network connection lost.","code":6001}],"success":false,"result":{},"messages":[]}

curl https://api.cloudflare.com/client/v4/accounts/.../ai/run/@cf/openai/whisper  \
  -X POST  \
  -H "Authorization: Bearer ..."  \
  --data-binary "@media.opus"

curl https://api.cloudflare.com/client/v4/accounts/.../ai/run/@cf/openai/whisper  \
  -X POST  \
  -H "Authorization: Bearer ..."  \
  --data-binary "@media.opus"

plgingras88•5/29/25, 8:58 PM

Unsure if it's all related but something seems odd with workers AI since yesterday. Nothing seems to have been raised in the status channel

SODaLC•5/30/25, 5:10 PM

Anyone know why AI responses keep switching between structed json and broken json strings? I am using @cf/meta/llama-3.2-11b-vision-instruct.

SSODaLC Anyone know why AI responses keep switching between structed json and broken jso...

Viktor•5/30/25, 8:38 PM

I noticed a lot of these errors when not using complete inputs; e.g. just sending a system prompt or not having a user prompt or tool response as last input.

VViktor I noticed a lot of these errors when not using complete inputs; e.g. just sendin...

SODaLC•5/31/25, 2:05 PM

Thats good to know for the future. I just checked to make sure; i did have a system prompt and user prompt in my input in that order.

Shados•6/1/25, 3:51 AM

Out of curiosity, what do folks who use the Cloudflare Worker AI stuff feel about its pricing model? I did some comparisons since we have Azure OpenAI and Cloudflare available at work, and unlike other CF offerings (eg: Cloudflare Workers themselves), it looked surprisingly expensive. The mid range llama models are priced about the same as Azure OpenAI Nano/Mini per million token, give or take (its not an apple to apple comparison), while I expected them to be much cheaper. The 10k neurons a day not withstanding (which, to be fair, does make a decent difference)

Is there something I'm overlooking? I'm just comparing input/output tokens here, so its a pretty naive comparisons, since again its tough to compare "equivalent" models between Llama models and the GPTs.

SShados 👋 Out of curiosity, what do folks who use the Cloudflare Worker AI stuff feel a...

zensin•6/1/25, 7:40 AM

I'm only using them for dev testing because of price and the models being offered are old/outdated. For me to consider using it, I would need up to date models and pricing similar to openrouter.

zensin•6/1/25, 7:48 AM

I expect there to maybe be a slight premium for being close to all the users and not having cold-start issues but the cases where that makes a big difference are few

zensin•6/1/25, 7:51 AM

the AI gateway on the other hand is great, and has a lot more potential to realize.

yinxingmaiming6409•6/1/25, 2:03 PM

Is there a way to obtain the image embedding on CF?

implicit-invocation•6/3/25, 3:27 AM

when to expect the new open source models like

gemini-3n

gemini-3n

and

chatterbox-tts

chatterbox-tts

? is there any place to request and vote?

octopus8717•6/3/25, 8:20 PM

Hello, I'm using worker with default LLM Chat App Template. Is it possible to connect remote MCP to it?

gabrielstuff•6/5/25, 6:09 PM

Hello, is there any roadmap for custom model support ? Like on the fly model download from hugging face ?

gabrielstuff•6/5/25, 6:19 PM

The only useful info I’ve come across so far is from here:
https://community.cloudflare.com/t/newer-ai-models-bring-your-own-model/597274
https://github.com/cloudflare/cloudflare-docs/issues/11039
https://developers.cloudflare.com/changelog/

At the moment, “bring your own model” is something very few platforms do well. Among the better ones, I’d say Scaleway Managed Inference and Deepinfra are leading the pack.

Being able to plug in your own model, especially for tasks like completion or classification, is a huge deal. It opens the door to running smaller, highly optimized models, including those trained for other languages or specific domains. And for fine-tuned models, this could really set Cloudflare apart if done right !

GitHub

Ai model deployment, can we deploy our models to cloudflare? · Iss...

Which Cloudflare product(s) does this pertain to? Workers Subject Matter AI workers Content Location https://github.com/cloudflare/cloudflare-docs/blob/production/content/workers-ai/models/embeddin...

Cloudflare Docs

Changelogs

Ggabrielstuff Hello, is there any roadmap for custom model support ? Like on the fly model dow...

Isaac McFadyen•6/5/25, 7:55 PM

?workers-ai-models

check readme

Similar Threads

Similar Threads

Similar Threads