Have you tried reading this? https://developers.cloudflare.com/vectorize/platform/limits/ I think ma

Have you tried reading this?
https://developers.cloudflare.com/vectorize/platform/limits/
I think main limit you might be concerned about would be the max upsert per batch

Cloudflare Docs

Limits

The following limits apply to accounts, indexes and vectors (as specified):

Yyinxingmaiming6409 Are you sure? How are workers restricted? Can you send it to me? I have never en...

U.Senthil Kumar | Contributor•6/14/25, 6:13 AM

Workers have some restrictions like fixed binding names (no dynamic binding names at runtime). You can’t dynamically create bindings per user, but you can handle routing logic inside the Worker code. Adding a teammate doesn’t multiply cost, pricing is based on usage, not the number of collaborators.

11984 Ford Laser Have you tried reading this? https://developers.cloudflare.com/vectorize/platfor...

U.Senthil Kumar | Contributor•6/14/25, 6:15 AM

Yes, the main limit you'd hit is the max upsert batch size: it's capped at 1,000 vectors per batch when using Workers, and 5,000 via the HTTP API, So if you're sending large batches from a Worker, you'll need to chunk them into batches of ≤ 1,000 vectors.Also note index name limits (≤ 64 bytes), but chunking your upserts is the key concern.

11984 Ford Laser Have you tried reading this? https://developers.cloudflare.com/vectorize/platfor...

yinxingmaiming6409•6/14/25, 7:18 AM

No, I'm not worried about data insertion restrictions, I'm mainly concerned about queries. Because I am working on an on-site service software that takes photos and checks for similar images. So there may be more queries.

yinxingmaiming6409•6/14/25, 11:39 AM

Does anyone know that if vectorize is operated by workers, there is no speed limit?

beingmudit [WePlayDOS]•6/15/25, 9:28 AM

Hey Team, is there a way to delete older vectors in a single go or if I want to clean up my entire index.

Bbeingmudit [WePlayDOS]Hey Team, is there a way to delete older vectors in a single go or if I want to ...

U.Senthil Kumar | Contributor•6/15/25, 1:06 PM

Yes, you can delete older vectors or clean up the entire index using a batch deletedelete operation or by dropping and recreating the index.

yinxingmaiming6409•6/16/25, 2:42 AM

I want to use queries extensively now, but who knows if workers really don't limit frequency like HTTP interfaces do?

yinxingmaiming6409•6/17/25, 1:18 AM

yinxingmaiming6409•6/17/25, 1:19 AM

Is this representing the number of times or the dimension of a vector?
For example, with 512 dimensions and 100 queries, will 51200 be displayed?

yinxingmaiming6409•6/17/25, 8:48 AM

After I store the image, I immediately retrieve the vector through the ID. However, he may not be able to retrieve it and may experience a delay of 2-3 seconds.

yinxingmaiming6409•6/18/25, 3:48 AM

What techniques should I use to cache the data read from the vector using a worker? What about KV?

yinxingmaiming6409•6/20/25, 2:18 PM

How can search directly filter scores? Do I have to search and filter by myself?

Stupefystar•6/23/25, 10:04 AM

I am getting 500 when I try and insert a new vector record: Error: VECTOR_INSERT_ERROR: Status + 500

yinxingmaiming6409•6/26/25, 8:02 AM

I hope to increase the ability to batch query the Id list in the vector, because sometimes my database data doesn't seem to correspond here. I deleted data from my database, but there is still data here.

yinxingmaiming6409•7/3/25, 4:28 AM

yinxingmaiming6409•7/3/25, 4:29 AM

The worker has a serious bug. I tested the output logs and found no issues with the data. However, the accuracy was lost after uploading.

yinxingmaiming6409•7/3/25, 4:29 AM

yinxingmaiming6409•7/3/25, 4:30 AM

Uploading through HTTP API will not lose accuracy

yinxingmaiming6409•7/3/25, 5:30 AM

Who can test uploading a float64 vector and then check if it will turn into float32?

yinxingmaiming6409•7/4/25, 2:46 AM

I can't just search by criteria, can I? For example, searching for all data with EId=14 in Metedata

wwb•7/4/25, 6:36 PM

I'm having a weird problem adding things to vectorize. I am using the typescript API from a worker, but I keep getting this error:
X [ERROR] Error: VECTOR_UPSERT_ERROR (code = 40023): vectorize.upstream_error - failed to parse upsert vectors request in ndjson format: line Some(0) was not expected format

wwb•7/4/25, 6:36 PM

which is strange as I am not really touching the json formatting at all, just handing it to VECTORIZE.upsert()

wwb•7/4/25, 8:01 PM

figured it out -- array subterfuge, you need embedding.data[0]embedding.data[0] not just embedding.dataembedding.data as your values.

yinxingmaiming6409•7/7/25, 9:02 AM

I found that the search and storage of float32 bits are the same, and their similarity is also more than 80%. I remember it was 100% before. Is your system unstable?

yinxingmaiming6409•7/7/25, 9:05 AM

It seems that when searching for the same vector, there were times when the score was 100%. Now there is a problem with the malfunction, and no matter how you search, it is still not 100%. The real-time storage and query vectors are identical. The data is exactly the same.

yinxingmaiming6409•7/7/25, 9:17 AM

Can anyone solve it?

yinxingmaiming6409•7/7/25, 10:04 AM

For example, for this vector, if you add it to the index and search for it using cosine method, the similarity is score=0.86116236, which is the same as the search

vector.txt6.23KB

yinxingmaiming6409•7/7/25, 10:04 AM

Yyinxingmaiming6409 It seems that when searching for the same vector, there were times when the scor...

yevgen•7/7/25, 10:10 AM

see https://developers.cloudflare.com/vectorize/best-practices/query-vectors/#control-over-scoring-precision-and-query-accuracy

Cloudflare Docs

Query vectors

Querying an index, or vector search, enables you to search an index by providing an input vector and returning the nearest vectors based on the configured distance metric.

Yyevgen see https://developers.cloudflare.com/vectorize/best-practices/query-vectors/#co...

yinxingmaiming6409•7/7/25, 10:11 AM

Isn't the property 'returnValues: true' controlling whether to return a vector or not?

yinxingmaiming6409•7/7/25, 10:11 AM

Is there also a function to control search accuracy?

yinxingmaiming6409•7/7/25, 10:12 AM

yinxingmaiming6409•7/7/25, 10:12 AM

I'm going to try it right now

yevgen•7/7/25, 10:12 AM

returnValues: true returnValues: true that's the only option to increase search accuracy. We decided against exposing a dedicated setting for accuracy.

yevgen•7/7/25, 10:13 AM

with returnValues: true returnValues: true you should reliably get >90% accuracy.

yinxingmaiming6409•7/7/25, 10:15 AM

The exact same data, can he definitely be 100%? Because this string can be judged 100%

yinxingmaiming6409•7/7/25, 10:17 AM

I have deployed a new worker now, and I will test it in a few minutes to see if it can truly query accuracy.

yinxingmaiming6409•7/7/25, 10:21 AM

I think it would be better to separately set whether to enable high-precision queries and whether to return values. Because I don't need to return values

benman•7/14/25, 5:39 PM

I had errors had to switch to the api

Abid•7/18/25, 11:39 PM

Hey, are there any plans to support arbitrary dimension counts? The documentation indicates that 1536 is the maximum. I am trying to use gemini-embedding-001 which can provide a 3072-wide vector and I would like to store the entire vector.

ldenoue•7/23/25, 7:46 AM

Is it possible to retrieve the vector of the centroids you compute over the vectors? I want to be able to use this to retrieve topics: once i have these centroid vectors, I can query the db for each and retrieve the corresponding top N chunks and ask an LLM to summarize these chunks.

yinxingmaiming6409•7/24/25, 2:07 AM

Querying is also charged based on dimensions. For example, if I query a 512-dimensional dataset and retrieve 10 pieces of data with 512 dimensions, should I be charged for 512 or 5120?

iLearn•7/24/25, 10:10 PM

Hi , Please help in acheiving my AIM, as explained below with reason.

AIM : want to have 1 Vector DB per User .. for this design, plan to have created 50000 Vectorize Bindings (Upper Limit by CF) to the Worker (in first deployment) .. then, as new user signs-up pick a deterministic allocation of 1 Binding from "available pool of bindings" ..

What after 50000 users signed up ? .. create a new CF account and somehow find a way to direct the traffic based on a KV Map which keeps track for Worker Account and Binding ID with-in that worker, allocated for the user.

Question : is there a smarter way to acheive the AIM ?

Reason for My Ugly Design : Upfront Heavy Cost Incurred, for "just ONE" query done .. AFTER "FIRST" query, the incremental costs for query are very minimal .. so, if there are many NAMESPACE seperated users in ONE Vectorise Index .. Just ONE Active user will baloon-up the bill for rest dormant users (who never queries DB in that month)
This will result in ONLY ONE PAID User, while Vectorize will costing me for ALL Users in that Index.

Example :
CF Pricing : ((queried vectors + stored vectors) * dimensions * ($0.01 / 1,000,000)) + (stored vectors * dimensions * ($0.05 / 100,000,000))
Assume : stored vectors = 4200000, dimension = 1024, duration = 1 month
Case-1 : ZERO QUERIES , Total Cost = $2.1504
Case-2 : THREE QUERIES, Total Cost = $45.15

IiLearn Hi , Please help in acheiving my AIM, as explained below with reason. AIM : wan...

Elliot Hesp•7/24/25, 10:11 PM

You'd be better using workers for platforms and dynamically binding a db

EElliot Hesp You'd be better using workers for platforms and dynamically binding a db

iLearn•7/24/25, 10:13 PM

Does Workers for Platforms have upper limits ? Can you please point to web-page which can help ? Thanks.

Elliot Hesp•7/24/25, 10:15 PM

Probably the same limits. You'd most likely be able to reach out to support to request an increase in max databases. Although if you have 5000 bound to a worker and they're not actually being used, it's unlikely to be a valid request

EElliot Hesp Probably the same limits. You'd most likely be able to reach out to support to r...

iLearn•7/24/25, 10:20 PM

Yeah true and also expected as such request will be an abuse. But, how to design per-user-DB .. as my architecture solves per-user problem beautifully with DO .. but, got stuck to assign those users private and "cost-effective" Vector DB ..

EElliot Hesp You'd be better using workers for platforms and dynamically binding a db

iLearn•7/24/25, 10:27 PM

Just checked with ChatGPT for this. Following is response from ChatGPT. Please confirm, if your were referring in some other context. Thanks.

The statement: "using Workers for Platforms you can dynamically bind a Vectorize DB per user, which otherwise is not possible using Workers Paid plan" is not accurate.

Cloudflare Workers—whether on Free, Paid, or Platforms—use static bindings defined in your wrangler.toml (or dashboard) configuration. You must specify each Vectorize index you want to connect to at build/deploy time, not dynamically at runtime

IiLearn Just checked with ChatGPT for this. Following is response from ChatGPT. Please c...

Isaac McFadyen•7/24/25, 10:53 PM

ChatGPT is incorrect. Workers for Platforms allows you to dynamically create Workers (and therefore bind Vectorize DBs to those Workers) via an API.

IIsaac McFadyen ChatGPT is incorrect. Workers for Platforms allows you to dynamically create Wor...

iLearn•7/25/25, 8:08 PM

Thanks for correcting the understanding.

Unfortunately, this also doesnot solve my AIM (= "Isolated private Vectorize DB "per User"), as i shall be hitting the limit
"Vectorize Indexes per account 50,000 (Workers Paid)"

Q#1 : How to scale after 50k (PROGRAMATICALLY) ?
Q#2 : IF, Q#1 is NOT possible, is Cloudflare Infra "capable" to handle Vectorize DB "per User", if manual request need to be dropped to support to increase the limits, as i start nearing the 50k limit ?

Thanks in advance. These Qs are very important for me to architect the whole AI pipeline with serverless approach for "planet earth". (DO solves one part by "storing chunks" for planet-earth, but Vectorize is bottle neck for full fledged AI pipeline).

Have you tried reading this? https://developers.cloudflare.com/vectorize/platform/limits/ I think ma

Similar Threads