azure / ai-hub Goto Github PK

Goal state

The AI hub provides the AI foundation which can be deployed to 1 or more Azure regions, adhering to strict security and compliance requirements our Enterprise customers have in place. On this foundation, we should explore how we can add additional value by leveraging other efforts to accelerate AI adoption and relevant use cases, such as the great example here: https://github.com/microsoft/sample-app-aoai-chatGPT

We should explore how we can provide the option to:

1. List and enumerate the deployable sample apps into the deployment experience
2. Explore how and what the 'contract' can be to ensure a deterministic and successful deployment
3. Implement the required telemetry for usage and adoption of the above.

Can we simplify the meta prompts and store them outside of the resource declaration, for easier maintenance and tweaking?

Goal state

Customer can select deployment option to have everything entirely on a private virtual network, including APIM.

Goal state

We need a detailed article that describes the various scenarios where APIM is adding value in context of Azure OpenAI.
Besides dynamic load balancing and retry/throttling logic etc., we must provide prescriptive guidance for known use-cases, such as:

• Combination of PTU and pay-as-you-go.
• Same as the above, where organisations can "start small & expand" without any dramatic changes required on the application side.
• Differen APIM policies to address different scenarios. We should ensure this is reflected and available as options in the reference implementation as we wrap things up.
• External vs Internal vs Hybrid
• Single region - and multiple regions

For the on-your-data scenario, when a document is analysed by Document Intelligence, the Azure Data Factory pipeline extracts the markdown content from the Document Intelligence assessment, but it saves the output as JSON with a single property: "content".

The pipeline should extract the markdown text from the "content" property, and save export the contents as a markdown file (.md)

Goal state

Provide an opinionated view and implementation of the e2e observability pattern that would help organisations to truly 'turn on the light' for their AI usage in Azure.

All services provides option to enable logging and threat detection (via Diagnostic Settings) as part of the deployment, and we should investigate the value-add by exploring:

• AzMonitor Workbook for the scenarios
• Curated alerts ant notification mechanisms of what is vital and critical
• Open to other creative ideas :-)

Currently, the on-your-data scenario, when multiple files need to be uploaded at the same time, the Azure Data Factory pipeline grabs one file at the time and triggers the on-your-data ingestion job.

It would be more efficient if the Azure Data Factory pipeline grabs all the files, places them in the blob container that is used by AI Search, and then trigger only once the ingestion job. The ingestion job will then vectorise and index all those files in a single job, rather than having to run that job multiple times, once per file.

Include the following enhancements for the On Your Data scenario within the reference implementation:

• In ADF, ensure output from Document Intelligence is preserved in markdown (and not in .txt) via a data flow.
• Make sure the following containers are created in the storage account: curated, docsbatch and docsintelbatch.
• Remove the docs and docsintel containers from storage account.
• In ADF, remove triggers for single file upload and single file upload with docIntel.
• In ADF, ensure AOAI On Your Data ingestion job is only executed one (for index and indexer creation).
• In ADF, implement pipelines for batchAnalyzeDocuments and batchAnalyzeDocuments_withoutIntel.
• Ensure triggers (based on Schedules) are implemented for the two pipelines described in the previous step.

Hi,

I've been attempting to deploy this, but disabled purge protection (solely for testing purposes), however the deployment always fails and doesn't get beyond the key vault deployment.

"code":"DeploymentFailed","message":"At least one resource deployment operation failed. Please list deployment operations for details. Please see
https://aka.ms/arm-deployment-operations
for usage details.","details":[{"code":"BadRequest","message":"{\r\n "error": {\r\n "code": "BadRequest",\r\n "message": "The property \"enablePurgeProtection\" cannot be set to false. Enabling the purge protection for a vault is an irreversible action."\r\n }\r\n}"}]}]}]}

Is this no longer an optional setting?

For the On-your-data scenario, update embeddings model:

• From: text-embedding-ada-002
• To: text-embedding-3-large

Add a filter to ensure that only supported file types are processed:

Azure Document Intelligence:

• PDF
• JPEG/JPG, PNG, BMP, TIFF, HEIF
• DOCX, XLSX, PPTX
• HTML

Azure OpenAI On Your Data

• TXT
• PDF
• DOCX
• PPTX
• MD
• HTML

As per comment in #131

@krnese krnese 7 hours ago
We need to add retrieval tool to the assistant

Member
@krnese krnese 6 hours ago
Line #146 using time.sleep within an orchestrator function is not recommended as it can cause unnecessary delays and increased costs. Instead, use create_timer for durable functions.

VideoIndexer-Workflow needs callback url from Callback-Workflow