Azure AI Foundry

Run inference against models deployed in Azure AI Foundry. Agents generate chat completions and structured output, create text and image embeddings for RAG pipelines, and inspect deployments through one unified endpoint across OpenAI, Mistral, Llama, and other providers.

4 actions API key available

Azure AI Foundry website ↗ Platform Documentation ↗ Capability data verified 2026-07-27

A new supplier contract lands in the intake mailbox and is routed to the flow

Create Text Embeddings vectorizes the document chunks for the retrieval index

Create Chat Completion extracts supplier name, payment terms, and renewal date as JSON

The flow validates the extracted JSON against the expected schema and flags low-confidence fields

An operations reviewer confirms or corrects the flagged fields before anything writes to the ERP

Create Chat Completion drafts a plain-language summary of the contract for the team

The summary posts to the team channel with a link back to the source document

What This Integration Enables

Azure AI Foundry is built for teams that standardized on Azure and refuse to marry a single model vendor. Deployments from OpenAI, Mistral, Llama, Phi, DeepSeek, and others sit behind one inference endpoint, and the model is just a deployment name you pass in. FlowRunner treats that endpoint as the reasoning step inside a governed workflow: agents call Create Chat Completion for extraction, drafting, and structured JSON output, Create Text Embeddings and Create Image Embeddings for retrieval pipelines, and Get Model Info to confirm a deployment exists before betting a workflow on it. And because FlowRunner is BYOK, the Azure subscription, the deployments, and the spend stay under your governance; the workflow layer adds the retries, approvals, and audit trail around them. The point is not that a model runs. It is that the model's output lands somewhere accountable: - Chat completions return structured JSON that downstream steps validate before acting on - Text and image embeddings feed vector stores like [Pinecone](/integrations/pinecone), [Qdrant](/integrations/qdrant), and [Azure AI Search](/integrations/azure-ai-search) - Tool and function calling lets a Foundry-deployed model drive other FlowRunner connectors - Deployment metadata checks catch a renamed or retired model before it breaks production flows - Every inference call is a logged workflow step, not an untracked script on someone's laptop

Without FlowRunner

One vendor, one SDK, one rewrite Every model switch means new client code, new auth, and a migration project

Extraction lives in scripts Document parsing logic sits in a notebook someone ran once and nobody governs

RAG is a side project Embedding pipelines run ad hoc, so the search index quietly drifts from the source documents

With FlowRunner

One deployment name, any provider Swapping GPT for Mistral or Llama is a config change in Foundry, not a flow rewrite

Prompts inside governed flows Extraction, summarization, and RAG steps run as auditable workflow nodes with retries and gates

Embeddings on a schedule New documents are vectorized as they arrive, and full re-indexes happen only with sign-off

Use Case Scenarios

The document intake line that never marries a model

Invoices and contracts arrive by email and are parsed by the flow. Create Chat Completion, pointed at the team's current extraction deployment, returns supplier, amounts, and dates as JSON. The flow validates the schema, writes clean rows to [Google Sheets](/integrations/google-sheets), and routes anything ambiguous to a reviewer. When the team wants to trial a cheaper model, they deploy it in Foundry under a new name and change one field in the flow. The workflow, the validation, and the audit trail do not move.

RAG over the policy library, kept honest

A nightly flow lists new and changed documents in [SharePoint](/integrations/sharepoint), chunks them, and calls Create Text Embeddings on each chunk. Vectors upsert into [Azure AI Search](/integrations/azure-ai-search) so the index tracks the source library instead of drifting from it. When an employee question arrives from [Slack](/integrations/slack), the flow embeds the question, retrieves matching chunks, and answers with Create Chat Completion, citing the documents it used. Questions that retrieval cannot ground get escalated to a person instead of answered anyway.

Cross-modal product search from one endpoint

A retailer's flow watches for new product images, pairs each with its title, and calls Create Image Embeddings against a multimodal deployment such as Cohere Embed v3. The vectors land in [Pinecone](/integrations/pinecone) alongside text embeddings of the catalog copy, so a support agent's query in plain language surfaces the right SKU even when the words never appear in the product description. Get Model Info runs at the start of each batch to confirm the embedding deployment still exists and still reports the expected type.

Human-in-Loop Highlight

The expensive mistake with Foundry is not a bad answer, it is a bad batch. Re-embedding an entire document corpus with Create Text Embeddings after a deployment swap is metered spend across millions of tokens, and it invalidates the existing vector index the moment the new vectors start landing, because embeddings from different models do not live in the same space. So when a FlowRunner agent detects that the embedding deployment behind the index has changed, it does not start the re-index. It posts the scope: how many documents, which deployment is being retired, which one replaces it, and what the token volume implies. A person approves the run, and only then does the corpus re-embed begin. The agent handles every chunk at machine speed; the decision to spend the tokens and rebuild the index belongs to someone who can answer for it.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

4 actions

Inference

Create Chat Completion Generates a model response for a chat-style conversation using any chat model deployed in your Foundry project. Supports sampling and length controls, stop sequences, tool and function calling, and JSON response formatting, and works uniformly across providers including OpenAI, DeepSeek, Mistral, Llama, and Phi. The workhorse for extraction, drafting, classification, and agent reasoning steps.

Embeddings

Create Text Embeddings Creates embedding vectors for one or more input strings using an embedding model deployed in your project. Returns a numeric vector per input for semantic search, clustering, classification, and retrieval-augmented generation, with optional dimensionality reduction on models that support it.
Create Image Embeddings Creates embedding vectors from images, optionally paired with text, using a multimodal embedding model such as Cohere Embed v3. Each input accepts a base64 data URL image with optional text and returns a vector for cross-modal search and similarity.

Deployments

Get Model Info Returns metadata about a model deployed under the inference endpoint, including its name, type, and provider. The preflight check that confirms a deployment exists and does what the flow assumes before inference runs against it.

Frequently Asked Questions

What can FlowRunner do with Azure AI Foundry?

FlowRunner agents can run Create Chat Completion, Create Text Embeddings, and Create Image Embeddings in Azure AI Foundry, plus 1 more action.

Does connecting Azure AI Foundry to FlowRunner require OAuth?

No. Azure AI Foundry connects to FlowRunner with an API key, no OAuth flow required.

Can Azure AI Foundry trigger a FlowRunner workflow automatically?

Azure AI Foundry doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with Azure AI Foundry

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us