search

Azure / GPT-RAG

Sharing the learning along the way we been gathering to enable Azure OpenAI at enterprise scale in a secure manner. GPT-RAG core is a Retrieval-Augmented Generation pattern running in Azure, using Azure Cognitive Search for retrieval and Azure OpenAI large language models to power ChatGPT-style and Q&A experiences.
https://azure.microsoft.com/en-us/products/cognitive-services/openai-service
MIT License
767 stars 125 forks source link

Azure AI Search - vector search with text-embedding-3-large not working #110

Open jucastag opened 4 months ago

jucastag commented 4 months ago

I created an index manually ingesting json's from blob storage. Json are simple text with 'contentVector' field with embeddings. I tried diferent dimensions for testing: 256, 1536, 2048 and 3072. While configuring the index 'contentVector' field It only allows me to use 2048 as maximum dimension and not 3072. Also.. while configuring the vectorizer it doesn't allow me to set my 'text-embedding-3-large' deployed model. Only text-embedding-ada-002' appears as an option. In terms.. the vector search through azure portal doesnt work with error InvalidVectorQuery: Query vector length must match the dimensionality of the field 'contentVector' in the vector field list"

framigni commented 4 months ago

Hi @jucastag related to your problem, we have tried to use a different Index as well, but unable to make the Orchestrator working with the new Index (see Issue #111 ) . Have you had a chance to test the RAG with your new Index ?

jucastag commented 4 months ago

@framigni yes.. we've tested the app with different indexes in the past. But we tested this by creating different indexes using new runs of setup.py of the ingestion code. Make sure the new created index has the same schema as the orchestrator is expecting to find.. specially the retireval functionality. Use log stream of the orchestrator function in azure portal to debugg. We are able to create vecttorizers... Are you sure you have the latest infra deployed? remember to set contentVector type of data as edm collection single when configuring the metadata fields of the index. Also the new infra uses different search and openai api version. that could also generate issues.

JakobStadlhuber commented 3 months ago

It's currently also not possible to create a plain new 'on your data' setup with the Azure OpenAI Studio its only possible to select the text-embedding-ada-002 model. i really hope this will change soon.