OpenAI API

Executes GPT-4 and GPT-5 models via REST API endpoints with support for streaming token-by-token responses using Server-Sent Events (SSE). Implements request batching, token counting via tiktoken library, and context window management up to 128K tokens. Handles both synchronous completion and asynchronous streaming patterns with automatic retry logic and rate-limit backoff.

Translates natural language descriptions into executable code using Codex models fine-tuned on public code repositories. Accepts code snippets, file context, and docstrings as input; returns syntactically valid code with language-specific formatting. Implements prompt engineering patterns for few-shot learning and chain-of-thought code generation. Supports 20+ programming languages with language detection and context-aware completion.

Processes images (JPEG, PNG, WebP, GIF) via the vision-enabled GPT-4 model to extract text, objects, spatial relationships, and semantic meaning. Accepts images as base64-encoded strings or HTTPS URLs; returns structured descriptions, OCR text, object detection results, and scene understanding. Implements multi-image comparison and visual question-answering patterns with support for high-resolution image analysis (up to 2048x2048 pixels).

Enables models to invoke external functions by generating structured JSON function calls based on natural language requests. Accepts OpenAPI-style JSON schemas defining available functions, parameters, and return types. Model generates function calls with arguments; client executes functions and returns results to model for final response generation. Implements automatic schema validation, parameter type coercion, and multi-turn function calling for complex workflows.

Converts text into high-dimensional vector embeddings (1536 dimensions for text-embedding-3-large) using transformer-based encoding. Accepts variable-length text inputs (up to 8191 tokens) and returns normalized vectors suitable for cosine similarity search. Implements batch processing for multiple texts in a single API call, reducing latency vs sequential requests. Embeddings are deterministic and compatible with vector databases (Pinecone, Weaviate, Milvus).

Maintains multi-turn conversation state by accepting a messages array with role-based context (system, user, assistant). Each message includes role, content, and optional metadata. Model processes entire conversation history to maintain context and coherence across turns. Implements automatic context window management, truncating older messages when approaching token limits. Supports system prompts for behavior specification and assistant-provided context injection.

Enables training of GPT-3.5-turbo and other models on custom datasets to adapt behavior for specific domains or tasks. Accepts JSONL-formatted training data with prompt-completion pairs or message-based examples. Implements supervised fine-tuning with automatic data validation, train/validation split, and hyperparameter optimization. Produces a new model checkpoint accessible via API with custom model naming. Supports batch evaluation and cost estimation before training.

Processes large volumes of requests asynchronously at 50% discount vs real-time API. Accepts JSONL file with multiple API requests; processes them in batches over 24 hours; returns results in JSONL output file. Implements request deduplication, automatic retry logic, and cost optimization. Suitable for non-time-sensitive workloads like data labeling, content generation, and analysis. Results are stored for 30 days.

Classifies text for policy violations including hate speech, violence, sexual content, and self-harm. Accepts text input and returns classification scores (0-1) for each category plus a binary flagged/not-flagged result. Implements multi-category detection with category-specific thresholds. Designed for content moderation workflows, user-generated content filtering, and safety guardrails. Processes requests with <100ms latency.