Thanks,

Aaron Kao

The biggest headache I’ve noticed: AI agents can misinterpret instructions, get stuck in loops, or fail when the input data isn’t perfectly formatted. Even when you think you’ve told it clearly what to do, small errors in prompts or context can cause unexpected results.

It got me thinking about how dependent we are on structured data and precise instructions when using AI agents, without that, the “automation” feels fragile and stressful.

I’m using the Make an API Call module within the Pinecone app, with the following relevant parameters:

• URL: /vectors/delete
• Method: POST
• Body:
  { “ids”: {{7.json}},
“namespace”: “apps” }

I’ve built a “knowledge” base UPSERT scenario in Make.com for a customer service voice agent. The scenario chunks document data before upserting into the database.

So dependent on the size, length of the document being upserted there can be any number of chunks.

Now if that document is amended for whatever reason, I wanted to first delete any relevant chunks associated with it before upserting the new updated document/information.

When I read through the API documents I got the impression that I should be able to delete multiple vectors using an array of ID’s in the body of the http request. As per the example I provided. [Page Not Found]

However, this didn’t work.

So at the moment if I want to delete vectors this has to be done one at a time.

In Make this uses a credit for each vector deletion.

It would be nice to just send a payload of vector id’s and delete the records in one go (one make credit)

Hope that helps…

Did you manage to solve this?

I’m interested in knowing more about your use case. We could extend the “Delete a Vector” module to support deleting multiple vectors (and the same for Upsert) but I’ll need an example to convince our dev team.

pasha.bitz:

When you first click on an index it has a “Search by Text” option but once you select any other option the “Search by Text” option is no longer available - see screenshot.

I also filed a support ticket as you requested.

This looks like a UI issue.

The option “Search by Text” appears at first, and then disappears from the very same index, as soon as another option is selected.

If you’re still seeing this unexpected behavior, could you widen the view of your screenshot so we can see the index configuration? Or perhaps open a support ticket so we can look into it more deeply with you.

I want to enable people to communicate with the models they create using RAG with the documents they upload. I am saving the files to a Supabase bucket. How can I transfer them to Pinecone? Does this process need to be done on the NextJS side? Most importantly, when using Pinecone, do I need to parse the files uploaded by users?

The token limits on Pinecone Assistant within free tier are cumulative and do not reset monthly. You can see more details at Pricing | Pinecone . For use beyond these limits, you would need to upgrade to Standard tier. The standard trial may be helpful if you need more credits for your testing or development. Thanks for using Assistant!

Adam
Engineering, Pinecone

First after change it’s look like Input token is reset to 0 , but today i checked and try to testing again with free tier , it’s showing my input token is 3.4 millions which is seem not to be possible in my understanding (I understand that Input token is reset every month.)

Why it’s been like this , can someone explain to me ? And most important questions , how I fixed it or I have to paid for Standard tier again because it’s not seem fair.

Thanks for reaching out to us! Yes, you can use our logo on the module description page, that’s not a problem. You can download it from our homepage by right-clicking on the logo in the upper left corner then choosing to save the brand assets.

I’d love to get the details of the integration added to our documentation, too. Would you mind finding a time that works for you so we can talk through that and get things moving? Calendly

In the meantime, take a peek at our documentation on attributing usage to your integration. That’ll be the first thing we want to get added so we can correctly gauge how many Drupal admins are adding Pinecone to their setup.

Cory

Might be useful to generally connect - we have a couple hundred production usage (from sites that do share, so probably higher) with your tool, perhaps could also go on your integrations page if you have a third-party integration area?

Here is the module homepage: https://www.drupal.org/project/ai_vdb_provider_pinecone

Here is the AI Core homepage https://www.drupal.org/project/ai (8K sites reporting usage of this but smaller percentage use RAG related + Pinecone competes with our Milvus, Opensearch, etc integrations of course).

ajohnston:

Then suddenly, with no further calls to Pinecone on our end, and with the same LSN still present

This sounds like maybe something to do with asynchronous requests getting processed later in your application?

Could you open a support ticket so we can look into this for you?

For examples:

Some tables header cols and others do not.

At some point that seems to have broken, and our integration code hasn’t really changed at all for this since we added it, at least not for this part. Consider the following psuedo-code:

target_lsn = upsert_all(…)
wait_for(target_lsn)

while True:
  results = query({ a filter matching the unique metadata of the vectors that were just upserted })
  print(results)

Particularly when a lot of vectors were uploaded, the first few queries return no results. We don’t use a top_k, so even nothing was even close to being a good match, it should have returned something. But the first few queries get nothing. Then suddenly, with no further calls to Pinecone on our end, and with the same LSN still present, we get results after a little bit, presumably after it’s actually caught up on freshness.

So, it seems like there’s an issue with freshness. No matter what, it should have gotten something from Pinecone, but it gets nothing. And then suddenly gets results, making me think the freshness LSN is inaccurate in some way. Any help would be appreciated, this is a considerable problem on our end, though we can mitigate it to an extent for now.

Other notes:

• We use a serverless index.
• We embed with OpenAI
• Paid account, no issues otherwise, no errors

Also, if you’d like to join the Pinecone community on Discord, there are others there that can help answer your questions in a little bit more real time. Invite: Pinecone Nearest Neighbors: Official Community

Do you have feedback on what would be a reasonable limit for your use case and your clients? I can share that with our product team.

Can you share some more details around how you would expect to use the exported backups?

Thanking you,
Varun

Note: this multimodal PDF feature is not yet supported in the Console yet.

We have someone looking closer at this from our end too, but that will help.

If this assistant thing is not the best way… what is the best way to upload these scanned PDFs and search for the text it grabbed?

Old, broken:

const filteredFiles = await assistant.listFiles({
    filter: { metadata: { external_file_name: '2023.md' } }
});

New, working:

const filteredFiles = await assistant.listFiles({
    filter: { external_file_name: '2023.md' },
});

or:

const filteredFiles = await assistant.listFiles({
    filter: { external_file_name: { '$ne' : '2023.md' } },
});

Give it a try and let us know how it goes.

Also, with asynchronous requests, it’s possible that your application is holding on to the operation for too long and it gets cancelled before it gets sent.

Please open a support ticket so we can look into this further for your specific organisation.