We're currently working on supporting local binding for vectorize, hopefully completed this month :s

We're currently working on supporting local binding for vectorize, hopefully completed this month :soontm:

JJerome We're currently working on supporting local binding for vectorize, hopefully com...

echoes221•10/5/24, 2:12 PM

That would be great, thank you. In the meantime I’m generating embeddings using workers ai, then storing them in pg vector / querying in Postgres, seems to work well enough!

uri•10/6/24, 8:48 AM

what is "local binding for vectorize" mean?

Uuri what is "local binding for vectorize" mean?

Ashley•10/6/24, 9:46 AM

For a lot of Cloudflare products, you can use a similar version of the product locally. So when you run “wrangler dev”, the binding still works, so you can run/test locally - for Vectorize, that isn’t available yet, so you have to run it remote to test it (either deploy, or wrangler dev remote)

crunchy•10/6/24, 11:08 PM

Hi is there an additional information on the eventual consistency behaviour, can I assume the vectors are available for search in order? does the mutation identifier have an meaning? can it be used as a high water mark? I have a use case were a want to group vectors, it's tricky to do without knowing when vectors will be availible? presumable without a hwm I'd have to use a queue with some kind of back off but that won't help if the vectors are not indexed in order. presumble then the only solution would be repair on each read but i'm capped to 100 nearest neighbours.

Ccrunchy Hi is there an additional information on the eventual consistency behaviour, can...

JeromeOP•10/7/24, 6:53 AM

Hi! The DB state is eventually consistent, and all mutations (upsert, insert, delete, create metadata index, ...) are processed in the strict order they were given to the API. This means the index state is always reflecting all the mutations that were given to the API up to the last applied mutation, processed in order.
The mutationId acts as a high watermark indeed, you can compare the mutationId returned by any mutation operation and the one you get by calling https://developers.cloudflare.com/vectorize/reference/client-api/#get-index-info ; this will return the vector count, the last applied mutationId (again, in the sequence ordered as provided to the API) and the last UTC datetime this mutation corresponds to (useful if you don't keep track of the mutationIds)

vvotekeb•10/10/24, 12:33 PM

Does anyone know of any methods to create a backup for an index? I'm looking for a script/code that can pull all the data from one index and insert it to another.

kingmesal•10/10/24, 11:30 PM

I'm very eager to move off of pinecone, purely from a simplicity perspective.

I can work around some limitations, however there are a couple big blockers.

Metadata filtering has to support more than =, !=... E.g. <, <=, >=, > ... Even a filter for in list or not in list.
Not a blocker, but having an actual view in the dashboard would be a big help
The comment I saw above about local Wrangler support removes the other blocker

Bryan•10/11/24, 11:46 AM

I can't understand why vectorize has so many limits. I don't think I can develop a production-level project under current limits.

Compared to Pinecone, which has ZERO limits as long as I pay for what I need.

Really hope one day all limits go away.

BBryan I can't understand why vectorize has so many limits. I don't think I can develop...

Isaac McFadyen•10/11/24, 2:54 PM

The Vectorize limits aren't artificial but are a result of the internal architecture/technology backing Vectorize. I expect they'll be raised over time (and in fact they were recently) but not lifted altogether.

Isaac McFadyen•10/11/24, 2:54 PM

And nothing has zero limits: Pinecone docs say they have them too, it just scales by plan (but there's an upper bound): https://docs.pinecone.io/reference/quotas-and-limits

Pinecone Docs

Quotas and limits - Pinecone Docs

Search through billions of items for similar matches to any object, in milliseconds. It's the next generation of search, an API call away.

IIsaac McFadyen And nothing has *zero* limits: Pinecone docs say they have them too, it just sca...

Bryan•10/11/24, 3:59 PM

Oh I missed this

IIsaac McFadyen The Vectorize limits aren't artificial but are a result of the internal architec...

Bryan•10/11/24, 4:03 PM

I know there must be some limits to protect the platform. But I think, at least, should the total records of the specific index be not limited?

I know I can separate the index manually, but I really hope the platform can auto scale and I don't need to care how many records I will insert.

BBryan I know there must be some limits to protect the platform. But I think, at least,...

Isaac McFadyen•10/11/24, 4:10 PM

It's not so much to "protect the platform" but more that the actual underlying storage engine has a limit.

zhawtof•10/13/24, 5:40 PM

Feedback:

One of our biggest fears for Vectorize is that all metadata filters have to be created at index creation.

Wondering if there are plans to allow additional metadata filters to be created after index creation.

Thanks

HHexMan I deleted my Vectorize DB and tried to recreate it with the same name but I get ...

Isaac McFadyen•10/13/24, 6:14 PM

Vectorize index names are current single-use even if deleted.

Isaac McFadyen•10/13/24, 6:15 PM

https://canary.discord.com/channels/595317990191398933/1152193114522525726/1288526039487610953

Mitya•10/13/24, 6:39 PM

Hi all, I'm new to Vectorize and have been working through this tutorial (request to documentation author: we could really do with a summary of what the code in step 6 does, after the code.)

We previously vectorised a note ("pepparoni is the best pizza topping"), and saved it to the Vectorize DB. I get all that. What I'm less clear on is step 6. I send in a question, presumably something like "what's the best pizza topping?", and then we retrieve the vectors relating to the note closest to that question.

So my question here is: what powers CF's ability to associate that question with that note? Is it because both contain the word pizza? If instead I'd asked "what is the best topping for a common Italian dish?", would it have still returned the vector (my note)?

Mitya•10/13/24, 6:42 PM

Right, so it would make that link even though I myself have only given it one note and one question? (i.e. this suggests preexisting knowledge)

Isaac McFadyen•10/13/24, 6:43 PM

As for if you have more than 1 vector though, yes, the model that generates the vectors understands association between topics.

Isaac McFadyen•10/13/24, 6:43 PM

If you use a proper embedding model it also understands actual text, not just words - i.e. it would understand your second case there.

Mitya•10/13/24, 6:44 PM

@cf/baai/bge-base-en-v1.5

@cf/baai/bge-base-en-v1.5

a proper embedding model?

Isaac McFadyen•10/13/24, 6:44 PM

Vectorize itself is just a DB that stores the vectors. You could throw any vector in there, not just word embeddings - I've used it for facial recognition embeddings, for example.

Isaac McFadyen•10/13/24, 6:44 PM

Yes.

Mitya•10/13/24, 6:45 PM

OK great. So it's not like, with just one vector in the DB, it would return that vector whatever my question was?

MMitya OK great. So it's not like, with just one vector in the DB, it would return that...

Isaac McFadyen•10/13/24, 6:45 PM

With one vector it would, yeah - that's what Leo was saying. You request the "top <n>" vectors, so it'll always return that 1 vector since you can't request "top 0"

Mitya•10/13/24, 6:45 PM

Ah, gotcha. So this becomes useful only once you've pumped it full of a lot of stuff?

MMitya Ah, gotcha. So this becomes useful only once you've pumped it full of a lot of s...

Isaac McFadyen•10/13/24, 6:46 PM

Depends on your use-case, but it will also return a distance.

Mitya•10/13/24, 6:46 PM

Otherwise my users asking "what's the capital of Mongolia" are going to be told about pizza

Isaac McFadyen•10/13/24, 6:46 PM

So if you have 1 vector about pizza, and the user asks about Mongolia, and the distance is huge you can return a preset "sorry, I don't know" response.

Isaac McFadyen•10/13/24, 6:46 PM

Whereas if the distance is small, the question is likely about pizza.

Mitya•10/13/24, 6:46 PM

OK! That makes sense. I don't think the tut discusses distance

Isaac McFadyen•10/13/24, 6:47 PM

Ah OK, yeah - I'm pretty sure it returns distances (or it did last I used it)

PRM•10/13/24, 6:47 PM

someone help me please

PRM•10/13/24, 6:48 PM

#cloudflare-ai

Original message was deleted

Mitya•10/13/24, 6:49 PM

Similarity = distance? The tut (step 6) seems to assume that if it retrieved a vector that it's fine to use with the LLM. No checks on distance/similarity, hence my question.

Mitya•10/13/24, 6:50 PM

Ah that's true. So distance is more important if there's no LLM in play i.e. you just want to return the note, I guess?

Mitya•10/13/24, 6:51 PM

Great, thanks everyone.

Isaac McFadyen•10/14/24, 4:05 AM

Also

distance = 1 - similarity

distance = 1 - similarity

similarity of 1 = perfect match
similarity of 0 = absolutely no match
distance of 0 = perfect match
distance of 1 = absolutely no match

Isaac McFadyen•10/14/24, 4:05 AM

(sorry for ping)

HHexMan I deleted my Vectorize DB and tried to recreate it with the same name but I get ...

yevgen•10/14/24, 10:23 AM

that should be addressed in the coming weeks.

Zzhawtof Feedback: One of our biggest fears for Vectorize is that all metadata filters h...

yevgen•10/14/24, 10:25 AM

It's possible to create metadata filters after index creation, though metadata search will not apply to already existing data.

Kkingmesal I'm very eager to move off of pinecone, purely from a simplicity perspective. ...

yevgen•10/14/24, 10:26 AM

re #2

an actual view in the dashboard would be a big help

an actual view in the dashboard would be a big help

Index analytics is available in dash

HHexMan Is there any way to add more than 1 namespace to a vector that I am inserting? ...

yevgen•10/14/24, 10:31 AM

You could use https://developers.cloudflare.com/vectorize/best-practices/insert-vectors/#metadata to add multiple fields describing vectors and index each of the metadata fields separately to support filtering.

Cloudflare Docs

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 efficient, even as new vectors are added or existing vectors updated.

Kkingmesal I'm very eager to move off of pinecone, purely from a simplicity perspective. ...

yevgen•10/14/24, 10:33 AM

Metadata filtering has to support <, <=, >=, > ... Even a filter for in list or not in list.

Could you please list the order of metadata filtering operations you need the most to get started?

IIsaac McFadyen And nothing has *zero* limits: Pinecone docs say they have them too, it just sca...

yevgen•10/14/24, 10:42 AM

That is so true regarding nothing has zero limits. We are happy to work with customers to increase most of the limits once they are close to hit them.

Yyevgen re #2 `an actual view in the dashboard would be a big help` Index analytics is a...

kingmesal•10/14/24, 1:38 PM

Fair point. Those have been there. I meant a way to view / query something from the dashboard, similar to KV, or D1. An abiltity to perform a sanity check. e.g. In pinecone I had a bug once where I was removing a bunch of vectors before inserting a batch up new records and I had a bug, it died before inserting the new records. In their dashboard I had the ability to see some basic in formation to know that the data was gone.

We're currently working on supporting local binding for vectorize, hopefully completed this month :s

Similar Threads

Similar Threads

Similar Threads