Prompt Flow

Enables users to define LLM application workflows as directed acyclic graphs using flow.dag.yaml files, where nodes represent tools (LLM calls, Python functions, custom code) and edges define data flow between them. The execution engine parses the YAML, validates node dependencies, and executes nodes in topological order with automatic input/output mapping. Supports prompt templating, variable interpolation, and conditional branching through node connections.

Allows developers to define flows as Python functions or classes decorated with @flow and @tool, providing programmatic flexibility for complex logic that doesn't fit DAG patterns. The framework introspects function signatures to extract inputs/outputs, manages dependency injection, and executes flows with full Python semantics including loops, conditionals, and exception handling. Supports both synchronous and asynchronous execution with automatic tracing integration.

Packages flows as REST API endpoints that can be deployed to various serving platforms (local Flask server, Azure Container Instances, Kubernetes, etc.). The framework generates OpenAPI schemas from flow inputs/outputs, handles request/response serialization, and manages flow lifecycle (loading, caching, cleanup). Supports both synchronous and asynchronous serving with automatic scaling on cloud platforms.

Provides command-line interface (pf command) and Python SDK for programmatic flow operations: creating flows, running flows, managing runs, executing evaluations, and deploying endpoints. The CLI supports both DAG and Flex flows, integrates with shell scripting for automation, and provides structured output (JSON) for parsing. The SDK exposes the same operations as Python classes for integration into larger automation systems.

Integrates with Azure ML workspaces for cloud-based flow execution, dataset management, and compute resource allocation. Flows can be registered in Azure ML, executed on managed compute (CPU, GPU clusters), and results stored in workspace. Supports Azure ML datasets, models, and environments for reproducible cloud execution. The promptflow-azure package handles authentication, workspace configuration, and resource management.

Enables creation of multiple prompt variants within a single flow, each with different templates, parameters, or LLM configurations. The framework supports variant selection at runtime (via input parameters or conditional logic), batch execution across variants, and metric comparison to identify best-performing variants. Variants are stored in the same flow definition with clear separation for version control.

Supports processing of images, PDFs, and other multimedia files within flows through built-in tools for image loading, document parsing, and content extraction. Flows can accept image inputs, pass them to vision-capable LLMs, and process extracted text. The framework handles file I/O, format conversion, and integration with LLM vision APIs (OpenAI Vision, Azure Computer Vision, etc.).

Defines a lightweight .prompty format (YAML frontmatter + Jinja2 template + optional Python code) that bundles prompt definition, configuration, and execution logic in a single file. The framework parses the frontmatter to extract model parameters (temperature, max_tokens), system/user message templates, and optional Python initialization code, then renders templates with provided variables and executes LLM calls. Enables version control of complete prompt artifacts without separate YAML/Python files.

Provides native tool nodes for calling LLMs (OpenAI, Azure OpenAI, Anthropic, Ollama, etc.) with unified interface abstracting provider-specific APIs. Each LLM tool accepts a prompt, model name, and parameters (temperature, max_tokens, top_p), handles authentication via connection objects, and returns structured responses with token counts. Supports streaming responses, function calling, and automatic retry logic with exponential backoff.

Enables developers to wrap Python functions as reusable tools via @tool decorator, which automatically generates JSON schemas from function signatures for input validation and documentation. The framework handles parameter type hints, docstrings, and optional metadata to create discoverable tools that can be referenced in DAG flows or composed in Flex flows. Supports tool versioning, input/output validation, and integration with the tool registry.

Processes multiple input records (from JSONL files or datasets) through a flow in batches, executing nodes in parallel across available CPU cores or distributed compute. The framework handles input chunking, node-level parallelization, result collection, and error handling per record. Supports both local batch execution and cloud-based distributed execution via Azure ML, with automatic result aggregation and failure tracking.

Provides a framework for defining evaluation flows that compute metrics (accuracy, F1, BLEU, custom metrics) on flow outputs by comparing against ground truth labels. Evaluation flows are themselves Prompt Flow flows that take predictions and labels as inputs and output metric scores. The framework aggregates metrics across batches, supports multiple evaluation runs, and enables comparison of flow variants to measure improvement.

Automatically captures execution traces for all flow runs, recording node inputs/outputs, LLM API calls, latency, and token usage. Integrates with OpenTelemetry for distributed tracing, exports traces to backends (Azure Monitor, Jaeger, etc.), and provides token counting across different LLM providers. The promptflow-tracing package instruments code at the decorator level to capture traces without explicit logging.

Tracks all flow executions as runs, storing inputs, outputs, execution logs, and traces in a local or cloud-based run store. Each run captures execution metadata (start/end time, duration, status), node-level outputs, and artifacts (generated files, images). Provides CLI and SDK interfaces to query runs, compare results across runs, and visualize execution traces in VS Code or web UI.

Provides a connection object abstraction for managing credentials (API keys, connection strings, OAuth tokens) used by tools and flows. Connections are stored securely (encrypted in local storage or Azure Key Vault), referenced by name in flows, and support multiple connection types (OpenAI, Azure OpenAI, SQL, HTTP, etc.). The framework handles credential injection at execution time without exposing secrets in flow definitions.