AI/ML6 min read

OpenAI API

Integrate GPT-4o, Whisper, DALL-E, and embeddings via the OpenAI API - covering chat completions, function calling, streaming, vision, and cost management.

OpenAIGPT-4APIFunction CallingStreaming

skill: OpenAI API category: AI/ML tags: [OpenAI, GPT-4, API, Function Calling, Streaming] readingTime: 6

The OpenAI API is the default starting point for AI applications - broad model selection, best tooling, and widest community. For privacy/on-premise requirements, consider running open-source models. For lower cost, consider Anthropic Claude or Google Gemini Flash. For embeddings specifically, text-embedding-3-small is among the best price/performance ratios.

Workflow

Set up the OpenAI clientThe official SDK handles auth, retries, and streaming - use it instead of raw HTTP calls.
Use function calling for structured outputFunction calling forces the model to output valid JSON matching a schema - reliable way to extract structured data from unstructured text.
Stream responses for real-time UXWithout streaming, users wait 5-10 seconds staring at a blank screen. Streaming shows the first token in ~200ms.
Analyze images with GPT-4o VisionGPT-4o accepts image URLs or base64 images - useful for document parsing, screenshot analysis, and visual QA.
Count tokens and manage costGPT-4o costs ~$5/M input tokens and $15/M output. Without tracking, a single bad prompt loop can cost hundreds of dollars.

Output Format

OpenAI API integration with structured output, streaming, vision, and token tracking.

client = OpenAI()
response = client.chat.completions.create(
    model='gpt-4o-mini',  # cheaper for prototyping
    messages=[system_msg, user_msg],
    max_tokens=1000,
    temperature=0.7,
)
result = response.choices[0].message.content

Examples

Input

Output

Edge Cases & Gotchas

Temperature 0 isn't deterministic - Temperature 0 is near-deterministic but not guaranteed due to floating-point non-determinism on GPUs. Don't rely on 100% reproducibility.
System prompt leakage - Users can extract system prompts with prompt injection. Don't put sensitive business logic or secrets in system prompts - they're not secure.
Context window vs. cost - Using GPT-4o with 100K token contexts costs 100x more than 1K contexts. Summarize long histories or use retrieval instead of expanding context.
JSON mode vs. function calling - response_format: {type: 'json_object'} forces JSON output but doesn't validate schema. Function calling validates against a specific schema - use it for structured extraction.

When to Break These Rules

Use Claude API (Anthropic) when tasks require very long documents (200K context) or the task benefits from Claude's stronger instruction following.
Use local models (Ollama, vLLM) when data privacy prevents sending content to external APIs.

Reference Files

https://platform.openai.com/docs/api-reference
https://platform.openai.com/docs/guides/function-calling