Cloudflare Developers•3y ago

Sure! Here's the example without the metadata.

embeddings.ndjson18.56MB

SSam Sure! Here's the example without the metadata.

Matt Silverlock•11/8/23, 1:57 PM

embeddingembedding isn't correct. It should be valuesvalues

SSam Sure! Here's the example without the metadata.

Matt Silverlock•11/8/23, 1:57 PM

https://developers.cloudflare.com/vectorize/learning/insert-vectors/

Insert vectors · Vectorize

Vectorize indexes allow you to insert vectors at any point: Vectorize will optimize the index behind the scenes to ensure that vector search remains …

SamOP•11/8/23, 2:04 PM

Yeah I also created a version with that and it still didn't work. I can't seem to generate the truncated (500 line ) version of the most recent file I uploaded because OpenAI deprecated their api. But here's the 1000 vector version.

embeddings.ndjson31.34MB

SamOP•11/8/23, 2:05 PM

I had a version of that as well and that didn't work. This would be truncated to 500 vectors.

SSam I had a version of that as well and that didn't work. This would be truncated to...

Matt Silverlock•11/8/23, 2:13 PM

Let us take a look.

MMatt Silverlock Let us take a look.

SamOP•11/8/23, 3:42 PM

It's the file attached to this latest message.

SSam Yeah I also created a version with that and it still didn't work. I can't seem t...

Matt Silverlock•11/8/23, 6:35 PM

We're pushing a bug fix in a bit: seeing an issue with the REST API (vs. the Workers API) that's preventing insert in some cases. Will reply back once deployed.

SamOP•11/8/23, 6:37 PM

Thanks!

SSam Thanks!

Matt Silverlock•11/9/23, 1:10 PM

This should be good now - I can import your file.

MMatt Silverlock This should be good now - I can import your file.

SamOP•11/9/23, 1:10 PM

Thanks!

Jerome•11/10/23, 2:40 PM

Hey there! Vectorize has been out a month now. We'd love your feedback if you have some that you haven't shared yet. What's good, what's bad, what could be better, we take it all.

huw•11/13/23, 4:47 AM

A question about metadata filtering—what will be the pros and cons for using the namespaces feature vs defining a namespace in metadata and filtering on that? Query speed?

(I’m looking at building a system that searches across multiple namespaces at once and interleaves their results)

Hhuw A question about metadata filtering—what will be the pros and cons for using the...

Old Man Maker•11/13/23, 10:21 AM

I'm sure Matt will clarify anything I get wrong here, but my Assumption (with a capital A) is that namespaces, much like the way they are implemented in KV for example, are a way for you to separate your vectorize instances entirely - e.g, you are binding to a namespace, and the vectorize instances you have in that namespace are available to query on.

Filtering - the difference would be that your metadata filtering is allowing you to reduce the size and number of vectors you are performing a query over, if that makes sense. I'm not sure if that would impact billing (because you are charged on the number of vector dimensions you query against as well as storage of said vectors) - I'm sure all of this will be announced along with the release of those changes, which hopefully are not too far away.

Jerome•11/13/23, 2:20 PM

@oldmanmeta @huw

namespaces [...] are a way for you to separate your vectorize instances entirely - e.g, you are binding to a namespace, and the vectorize instances you have in that namespace are available to query on.

Jerome•11/13/23, 2:23 PM

spot on; also metadata filtering isn't available yet, so namespace filtering is the only filtering besides vector query we have right now.
Once metadata filtering lands, filtering on namespace and filtering on a namespace property in metadata will be functionally identical, although namespace filtering should induce lower latency due to the fact that we're able to index it in a way that makes it faster overall.

JJerome spot on; also metadata filtering isn't available yet, so namespace filtering is ...

huw•11/13/23, 9:50 PM

Exactly the answer I was looking for, thank you! I think for my use case it’s gonna make sense to use metadata filtering long-term

wadefletch•11/14/23, 5:26 AM

Where are the docs for namespace filtering?

Wwadefletch Where are the docs for namespace filtering?

Matt Silverlock•11/14/23, 1:38 PM

namespaces are part of the TypeScript/JS API; we haven't documented them yet as there was a bug when we initially launched. On our TODO list.

steyblind•11/16/23, 3:58 PM

OK, so I'm play-testing this, and I'm only a few weeks in to working with all this the fun offerings here, so looking for guidance based on what's coming (or not):

Goal:
- multi-tenant RAG CX via Slack
- support alpha concept (2-3 potential customers)

Context:
- I'm all set with workers, queues, data sync, AI gateway, etc.
- sync'ing data from Hubspot, Slack, Gong.io ~2.5k records per customer, maybe? Would expect this to be ~10k-25k+ in a production setting w/long-lived customer. Or more.

Experiment:
- How can Vectorize support this? (rip & replace pinecone)

System Design:
- Vectorize: id: (hex sha1 of text), values: [vector of text], metadata: {text, companyId}id: (hex sha1 of text), values: [vector of text], metadata: {text, companyId}
- D1 (possibly KV instead?): proxy for clerk.com Organization object + maps Slack IDs to users to Companies

Usage (simplified):
utterance from Slack

get company ID

query vectorize for context

return to response builder

bounce off LLM + augmented context

respond to Slack

So now what?
At this point I'm just waiting for filters or namespaces I guess, before this is "usable"?

Tumbzilla•11/16/23, 4:08 PM

Getting the following errors fairly frequently.

VECTOR_QUERY_ERROR (code = 2006)
VECTOR_QUERY_ERROR (code = 1000)

Couldn't find any docs on how to resolve. Anyone know of any docs that might point me in the right direction?

TTumbzilla Getting the following errors fairly frequently. VECTOR_QUERY_ERROR (code = 200...

Matt Silverlock•11/16/23, 5:54 PM

Is this just on a query?

MMatt Silverlock Is this just on a query?

Tumbzilla•11/16/23, 6:15 PM

yep. Just on queries. Roughly 1/3 of my requests are failing this way.

Tumbzilla•11/16/23, 6:16 PM

most of them are the first error code (2006)

TTumbzilla yep. Just on queries. Roughly 1/3 of my requests are failing this way.

Matt Silverlock•11/16/23, 6:17 PM

Can you share your index ID here?

MMatt Silverlock Can you share your index ID here?

Tumbzilla•11/16/23, 6:27 PM

just DM'ed it to you.

Hhuw Here’s what I have 🙂 ```toml vectorize = [ { binding = "embeddings-developmen...

JuanWTF•11/16/23, 10:44 PM

Just to confirm, is this the right way to bind many indexes to a worker?

JJuanWTF Just to confirm, is this the right way to bind many indexes to a worker?

huw•11/16/23, 10:51 PM

Yep!

huw•11/16/23, 10:52 PM

Multiple indices vs multiple namespaces will depend on what you’re trying to achieve

JuanWTF•11/16/23, 10:59 PM

What would be multiple namespace in this context?

MMatt Silverlock Can you share your index ID here?

Tumbzilla•11/17/23, 2:10 PM

hey @Matt - Any luck identifying the cause of the issue? Is the problem something I'm doing on my side?

Jerome•11/17/23, 9:39 PM

@Tumbzilla DM'd you

JJuanWTF What would be multiple namespace in this context?

huw•11/18/23, 1:38 AM

Using the namespacenamespace option when inserting & querying vectors (undocumented beta feature)

huw•11/18/23, 1:38 AM

It’s just a simple way to segment your index

GuillaumeLakano•11/18/23, 9:20 AM

Hi! How is priced getByIds() calls please? It's not really a querying, and from what I understood the Pricing is based on stored content and querying, but not about the operation that insert inside Vectorise.

GGuillaumeLakano Hi! How is priced getByIds() calls please? It's not really a querying, and from ...

Matt Silverlock•11/18/23, 4:13 PM

It is still a query. You are just scoping it to specific IDs.

scotto•11/18/23, 4:31 PM

metadata filtering will be announced here? any ETA?

steyblind•11/19/23, 12:23 AM

what bit do I need to flip to pick up the latest types?

steyblind•11/19/23, 12:23 AM

compatibility_date = "2023-11-08"compatibility_date = "2023-11-08"

kian•11/19/23, 12:47 AM

There hasn’t been a workerd/types release since that functionality was added

kian•11/19/23, 12:57 AM

At a glance, I’m also not sure if the flag is actually available yet.

kian•11/19/23, 12:57 AM

I’m not at a device to test, but the last attempt I can see shows No such compatibility flag: vectorize_query_metadata_optionalNo such compatibility flag: vectorize_query_metadata_optional

steyblind•11/19/23, 2:24 AM

Gotcha. Saw it here https://developers.cloudflare.com/vectorize/learning/query-vectors/

Query vectors · Vectorize

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 …

GuillaumeLakano•11/19/23, 2:28 PM

Hi, for our project, we need to known how we can structure our vectors, as currently in beta it's limited to 200,000 vectors / index.
Technically we only need one index, but ideally unlimited vectors.
Could you tell us if we can expect at least thousand of millions vectors / index in GA, or if we need right now to split our vectors inside multiple indexes, and queries each individual indexes to have our final results ?

GGuillaumeLakano Hi, for our project, we need to known how we can structure our vectors, as curre...

Matt Silverlock•11/19/23, 8:50 PM

vectorize

GGuillaumeLakano Hi, for our project, we need to known how we can structure our vectors, as curre...

Matt Silverlock•11/19/23, 8:51 PM

If you need “thousands of millions” (billions) per index, then you should absolutely split across multiple indexes. Running a 1B+ vector index on Pinecone or elsewhere would also be $$$$