@azure/ai-projects

Provides TypeScript/JavaScript SDK for initializing authenticated clients to Azure AI Projects service using Azure SDK credential chain (DefaultAzureCredential, ClientSecretCredential, etc.). Handles token refresh, credential fallback, and multi-environment authentication (cloud, sovereign, custom endpoints) through a unified client factory pattern that abstracts Azure authentication complexity.

Enables declarative configuration and deployment of AI models (LLMs, embeddings, vision models) to Azure AI Projects with model registry integration, endpoint management, and inference parameter specification. Abstracts model versioning, compute allocation, and deployment orchestration through a fluent API that maps to Azure's underlying model deployment infrastructure.

Enables models to return structured outputs (JSON, objects) that conform to a specified JSON Schema, with automatic validation and parsing. Defines response schemas declaratively, and the SDK ensures model outputs match the schema before returning to the application. Supports complex nested schemas, enums, and conditional fields with detailed validation error messages.

Supports passing multiple input modalities (text, images, PDFs, documents) to vision-capable models with automatic format conversion and preprocessing. Handles image encoding, document OCR, and multi-page document chunking transparently, allowing developers to pass raw files and have the SDK prepare them for model consumption. Integrates with Azure Document Intelligence for advanced document understanding.

Provides built-in rate limiting and quota management to prevent exceeding Azure API limits and manage token budgets. Implements token bucket algorithm for rate limiting, tracks quota usage across requests, and provides warnings when approaching limits. Supports configurable rate limits per model and automatic request queuing when limits are exceeded.

Provides a framework for building AI agents that can invoke external tools and APIs through structured function calling. Implements schema-based tool registration, automatic parameter binding, and execution result routing back to the model, supporting multi-turn agentic loops with state management across turns. Integrates with Azure AI Projects' native agent runtime for serverless execution.

Provides a centralized prompt registry within Azure AI Projects for storing, versioning, and retrieving prompts with variable substitution support. Enables teams to manage prompts separately from application code, with version history, rollback capabilities, and metadata tagging. Prompts are stored server-side and retrieved via SDK, supporting A/B testing and gradual rollout of prompt changes.

Provides SDK support for running evaluations against AI model outputs using built-in or custom evaluators, collecting metrics (accuracy, latency, cost), and storing results for analysis. Integrates with Azure AI Projects' evaluation runtime to execute evaluators at scale, supporting batch evaluation of large datasets and real-time monitoring of production model outputs.

Provides automatic instrumentation of AI application execution (model calls, tool invocations, agent steps) with distributed tracing support. Captures execution traces with timing, token usage, costs, and errors, storing them in Azure for analysis and debugging. Integrates with OpenTelemetry for standards-based observability and supports custom span creation for application-specific instrumentation.

Provides SDK-level content filtering for AI model inputs and outputs using Azure's safety policies. Filters prompts and completions against configurable safety categories (hate, violence, sexual, self-harm) with configurable severity thresholds. Integrates with Azure AI Projects' safety infrastructure to enforce organizational policies consistently across all AI applications.

Provides utilities for managing multi-turn conversation state including message history, context window optimization, and token counting. Automatically tracks conversation history with role-based message formatting (user, assistant, system), handles context truncation when exceeding model limits, and provides token counting to estimate costs before API calls. Supports conversation persistence to external storage with serialization/deserialization.

Enables asynchronous batch processing of multiple inference requests to Azure AI models with cost optimization and throughput maximization. Submits batches of prompts for processing, polls for completion status, and retrieves results with automatic retry and error handling. Supports cost-optimized batch APIs for non-latency-sensitive workloads, reducing per-token costs by 50% compared to standard inference.

Provides SDK support for generating vector embeddings from text using Azure's embedding models and storing them in integrated vector databases. Handles embedding model selection, batch embedding generation, and integration with Azure Cognitive Search or other vector stores for semantic search and RAG applications. Supports multiple embedding models with different dimensionality and performance characteristics.