Hi!
I am attempting to upload multiple documents at once in to a Retool AI Vector.
I am using a workflow to loop through rows in a retool database and upload each one as a document in to a vector.
This process works successfully when my Retool Vector uses either " text-embedding-2-small" or "text-embedding-ada-002". But it does not work when i try to upload a document using "text-embedding-3-large" as the embedding.
This is the error I see: Error: expected 3072 dimensions, not 1536
@Nathan_Franklin I was able to repro the error on my end. One thing I noticed is that it works fine when using the "Upsert Document" action, but fails when using the "Insert Document" action.
We are working on a fix, but would you be able to try changing the action type to "Upsert Document" as a temporary workaround in the mean time?
Also, @nicos this is a separate issue (should I make a new post?), but when making Retool AI resource queries using the different vectors which are both the exact same documents, the version referencing text-embedding-3-small vector is failing to find any relevant information and documents for almost any query while when i reference the vector using text-embedding-ada-002 it does quite well.
Screenshots show same query, same documents, very different performance. and in theory 2-small should outperform ada-002 since it's newer and better
@Nathan_Franklin would you be able to double check if the documents you are expecting are in the vector? (You should be able to see the listed documents in the UI if you navigate to the /vectors page)
Yep - for both vectors I see the same documents actually.
Actually for all three vectors I have the same documents in them.
And it works pretty much identically for queries using Retool Vectors text-embedding-3-large or text-embedding-ada-002 but not really at all for my queries using Retool Vector text-embedding-3-small.
Again same documents in all 3
Thanks for sharing that additional info @Nathan_Franklin. I am still having trouble repro-ing on my side, so I have a couple more questions to help debug further.
Are you using tags/metadata on any of your Vectors? (These can act as filters)
Do you see similar results for each of the embeddings models when using the "Similarity Search" Vector action?