Azure OpenAI
Put Azure OpenAI to work on your own records, inside your rules, with a person approving what it changes.
27 ways to read, 10 ways to write, each write switched on by you.
Quality checks
Built and tested with example responses.
- Ready to connect: doneEvery listed action is built in.
- Built and tested with example responses: doneEach action is built from the provider’s published API and tested against example responses, errors, rate limits and paging.
Connected, approved, proven.
The things teams do most with Azure OpenAI, an example flow with a person in the loop, and what Nexra keeps as proof.
Top actions
- Create batchApproval by rule
- Add files to vector store in a batchApproval by rule
- List batchesRead
Example flow
- Azure OpenAI · List batchesDone
- Azure OpenAI · Create batchNeeds approval
- Nexra · Record the decision and result in the evidence chainProven
Example workflow. Routine runs go through on their own. Only a change your rules flag, such as one over a limit you set, waits for the right person, who sees the exact change and the rule before it reaches Azure OpenAI.
What gets logged
- Whether it ran on its own or who approved it, and under which rule
- A fingerprint of the exact request sent to Azure OpenAI
- Azure OpenAI's reply: what it created or changed
- One entry in the evidence chain, linked to the one before
Try this in the sandbox
Run the whole flow with test data in your sandbox, then connect your own account when you are ready to go live.
For developers
Technical detail
Chat, Responses and embeddings on your Azure OpenAI deployments, plus files, batches, fine-tuning, vector stores, evals and containers on the GA v1 data plane.
Microsoft API docs- Source actions
- 27
- Destination actions
- 10
- Data types
- 19
What you can map
Reads are always available. Writes are off until you switch them on for a connection, and governed writes follow your policies.
Batches3
- Create batchDestinationNot retried automaticallyApproval before production
Starts a 24-hour batch from an uploaded JSONL file (purpose batch) against one endpoint. Batch requests are billed at the batch rate.
- Get batchSource
Status, request counts, output file and usage for one batch.
- List batchesSource
Batch jobs in the project, newest first, with request counts and file IDs.
Chat completions1
- Generate chat completionDestinationNot retried automaticallyApproval before production
Chat completion from a deployment. Supply a JSON schema for Structured Outputs.
Container files2
- Get container fileSource
Metadata for one file in a container (no contents).
- List container filesSource
File metadata in one container (no contents).
Containers2
- Get containerSource
One code interpreter container.
- List containersSource
Code interpreter containers with status, memory limit and expiry.
Conversations1
- Get conversationSource
Metadata of one Responses API conversation (items are not read).
Embeddings1
- Create embeddingsDestinationNot retried automaticallyApproval before production
Embeds up to 2,048 texts with an embedding deployment; one record per text.
Eval runs2
- Get evaluation runSource
Status, results and usage of one evaluation run.
- List evaluation runsSource
Runs of one evaluation with status, pass and fail counts and token usage.
Evals2
- Get evaluationSource
One evaluation definition (grader prompts are not returned).
- List evaluationsSource
Evaluations with data source type and grader names.
Files2
- Get file metadataSource
Metadata for one file (no contents).
- List filesSource
File metadata in the project (no file contents).
Fine tuning checkpoint permissions1
- List fine-tuned checkpoint permissionsSource
Projects a fine-tuned model checkpoint is shared with.
Fine tuning checkpoints1
- List fine-tuning checkpointsSource
Checkpoints saved during one fine-tuning job, with step metrics.
Fine tuning events1
- List fine-tuning job eventsSource
Status and metrics events for one fine-tuning job, newest first.
Fine tuning jobs2
- Get fine-tuning jobSource
Status, hyperparameters, trained tokens and result files for one job.
- List fine-tuning jobsSource
Fine-tuning jobs in the project with status, base model and resulting model.
Models2
- Get modelSource
One model available to the resource.
- List modelsSource
Models available to the resource (GA v1 API).
Responses2
- Generate response (Responses API)DestinationNot retried automaticallyApproval before production
Generates a reply with the Responses API, which Microsoft recommends for Azure OpenAI models.
- Get responseSource
Reads a stored response by ID.
Vector store file batch files1
- List files in a vector store file batchSource
Files attached by one vector store file batch, with processing status.
Vector store file batches2
- Add files to vector store in a batchDestinationNot retried automaticallyApproval before production
Attaches up to 500 uploaded files to a vector store in one batch (billable storage).
- Get vector store file batchSource
Progress of a vector store file batch.
Vector store files4
- Add file to vector storeDestinationNot retried automaticallyApproval before production
Attaches an uploaded file to a vector store; Azure OpenAI chunks and embeds it (billable storage).
- Get vector store fileSource
Ingestion status, usage and attributes for one file in a vector store.
- List vector store filesSource
Files attached to a vector store with ingestion status and attributes.
- Update vector store file attributesDestinationProvider duplicate protectionApproval before production
Replaces the filterable attributes of a file in a vector store.
Vector stores5
- Create vector storeDestinationNot retried automaticallyApproval before production
Creates a vector store, optionally with already-uploaded files. Storage above the free allowance is billed per GB per day.
- Get vector storeSource
One vector store with status, file counts and expiry.
- List vector storesSource
Vector stores in the project with file counts and storage usage.
- Search vector storeDestinationNot retried automaticallyApproval before production
Semantic search over a vector store's chunks with optional attribute filters; one record per matching chunk group.
- Update vector storeDestinationProvider duplicate protectionApproval before production
Renames a vector store or changes its expiry policy or metadata.
Access and limits
- Authentication
- Azure OpenAI resource key (api-key header)
- Plans and access
- An Azure subscription with an Azure OpenAI or Microsoft Foundry resource and model deployments. Calls use deployment names, not model names. Quota is assigned per deployment in tokens per minute.
- Test environment
- Use a low-quota deployment in a non-production resource for testing.
- Rate limits
- Tokens-per-minute and requests-per-minute quota per deployment; HTTP 429 with Retry-After when exceeded.
- Pagination
- Model lists return a single page; files, batches, fine-tuning jobs, vector stores, evals and containers page with after cursors (last_id) while has_more is true.
- Provider events
- Azure OpenAI inference does not deliver webhooks. Flows run on demand, on a schedule, or when a scheduled check finds new or changed records.
Before you build a flow
- Deployments cannot be listed with a resource key; that requires the Azure Resource Manager API and Microsoft Entra ID. Enter deployment names directly.
- The v1 models list returns models available to the resource, not deployments.
- Streaming, tools, images, audio and realtime are not exposed. The Assistants API (threads, messages, runs) is deprecated in favour of the Responses API and is not implemented.
- Files, batches, fine-tuning jobs and vector stores use the GA v1 data plane (/openai/v1). File upload (multipart) and file content downloads are not implemented; batch input files must be uploaded outside Nexra.
- Batch creation needs a Global Batch or Data Zone Batch deployment and supports /v1/chat/completions and /v1/embeddings; Azure's v1 schema does not list batch metadata, so it is not sent.
- Batch creation, vector store creation, vector store file attachment and vector store search are billable and never retried automatically.
- Fine-tuning jobs, events and checkpoints are read only; creating, pausing, resuming, cancelling and copying jobs are not implemented (fine-tuning is expensive and long-running). Checkpoint copy status needs a preview header and is not implemented.
- Evaluations, evaluation runs, containers, container files and conversations are read as metadata only: grader prompts, data source templates, run output items, container file contents and conversation items are never mapped. Creating, running, cancelling or deleting them is intentionally excluded.
- Deleting files, vector stores, vector store files and responses is intentionally excluded.
- Microsoft Entra ID tokens are not supported; use a resource key.
- Generation and embeddings are billed; they must be enabled per connection and are never retried automatically.
Developer reference
- Create the Microsoft credentials described above with the narrowest permissions your flows need.
- In Nexra, open Connect, then Connections, then New connection, choose Azure OpenAI, enter its settings and credentials, and tick only the write actions you need. Connection guide
- Test the connection, then use its actions as flow sources and destinations. Flow guide
Machine-readable detail: GET /api/public/connect/connectors/azure-openai and the Connections API.