This is AMAZING! Initial thoughts 💭 1- How much data can be embedded? 2 - Does response quality decline linearly with embedded database size? 3 - Can this handle multidimensional Daya (xls/CSV)? 🤔 All thoughts/answers/responses highly appreciated. 🤜🤛
Hi @nuclear_AI! Somewhat long response, but hope this helps provide some clarity 1 - amount of data in your vector store varies by data source you use, and the pricing tier of your search resource. For example, if you use AI Search in the Basic tier, you have 15GB of storage and 3 partitions (5GB each) with approximately 1.1 billion floats per partition. A lot of that is determined by how your data is chunked - larger chunks mean less embedded vectors, however smaller chunks offer more granular (and accurate) responses (but take up more space) 2 - response quality doesn't decline with embedded database size. Since all of your grounding data is stored as individual vectors in your database, the input is then vectorized and compared to nearby vectors - all part of the magic providing semantic results. There is possibility with too much data you'll have too many vectors in a small multi-dimensional area, and since the top responses are what are returned by the search resource, the data you *actually* want could get missed if the user input isn't specific enough. Again, this can vary greatly depending on how your data is cracked and chunked when writing to your vectorized index. 3 - Similar to standard indexes, how you delimit files like CSV/xls plays a role in how well that data is represented in the vector index. You can specify the delimiter in a custom index to split it however makes most sense for your data. When working with these multidimensional files, you'll likely need to test and iterate on your chunking methods to determine which way to split the files works the best. There is a great tutorial on the docs page that generates embeddings from a CSV that you might find helpful: learn.microsoft.com/en-us/azure/ai-services/openai/tutorials/embeddings?tabs=python-new%2Ccommand-line&pivots=programming-language-python
Hi there, I'm quite new to Microsoft and trying to understand more about the AI capabilities. Can we create copilots that work alongside Azure Dev Ops. For example i'd like to create a copilot that can generate user stories and then create those tickets on my board. I'm also really interested in creating a copilot that specialises in test generation, would it be possible to set a copilot up so that it reads my repo as a data source then generate e2e tests for it. Maybe it even outputs the tests in a new branch back on the repo?
Misleading title as you have NOT built a Copilot. You just played within the Azure AI Studio playground. Where is the Copilot that can be used by users?
This is AMAZING!
Initial thoughts 💭
1- How much data can be embedded?
2 - Does response quality decline linearly with embedded database size?
3 - Can this handle multidimensional Daya (xls/CSV)?
🤔
All thoughts/answers/responses highly appreciated.
🤜🤛
Hi @nuclear_AI! Somewhat long response, but hope this helps provide some clarity
1 - amount of data in your vector store varies by data source you use, and the pricing tier of your search resource. For example, if you use AI Search in the Basic tier, you have 15GB of storage and 3 partitions (5GB each) with approximately 1.1 billion floats per partition. A lot of that is determined by how your data is chunked - larger chunks mean less embedded vectors, however smaller chunks offer more granular (and accurate) responses (but take up more space)
2 - response quality doesn't decline with embedded database size. Since all of your grounding data is stored as individual vectors in your database, the input is then vectorized and compared to nearby vectors - all part of the magic providing semantic results.
There is possibility with too much data you'll have too many vectors in a small multi-dimensional area, and since the top responses are what are returned by the search resource, the data you *actually* want could get missed if the user input isn't specific enough. Again, this can vary greatly depending on how your data is cracked and chunked when writing to your vectorized index.
3 - Similar to standard indexes, how you delimit files like CSV/xls plays a role in how well that data is represented in the vector index. You can specify the delimiter in a custom index to split it however makes most sense for your data. When working with these multidimensional files, you'll likely need to test and iterate on your chunking methods to determine which way to split the files works the best. There is a great tutorial on the docs page that generates embeddings from a CSV that you might find helpful: learn.microsoft.com/en-us/azure/ai-services/openai/tutorials/embeddings?tabs=python-new%2Ccommand-line&pivots=programming-language-python
Thanks Folks that was nice and quick
Hi there, I'm quite new to Microsoft and trying to understand more about the AI capabilities. Can we create copilots that work alongside Azure Dev Ops. For example i'd like to create a copilot that can generate user stories and then create those tickets on my board. I'm also really interested in creating a copilot that specialises in test generation, would it be possible to set a copilot up so that it reads my repo as a data source then generate e2e tests for it. Maybe it even outputs the tests in a new branch back on the repo?
Great questions! I have the same ones and looking for answers. Have you found any information on this yourself yet?
you have an endpoint... then...? can connect to copilot studio?
❤ Microsoft learn
Misleading title as you have NOT built a Copilot. You just played within the Azure AI Studio playground.
Where is the Copilot that can be used by users?
🧠🧠