IntermediateИнженерия

OpenAI Cookbook: Official Recipes

Official OpenAI examples: Assistants API, function calling, embeddings, meta-prompting, vision, hallucination guardrails, data extraction, and fine-tuning. Real Python code from the OpenAI team.

8modules

8lessons

200 mintotal time

Developers learning advanced OpenAI API patternsaudience

Module 1

Assistants API and Agents

Building persistent assistants with tools: Code Interpreter, File Search, and custom function calls via the Assistants API.

Assistants API: Building Stateful Assistants

Deep dive into the Assistants API as a stateful evolution of Chat Completions: Thread, Run, Message primitives, built-in Code Interpreter and File Search tools, and run lifecycle management.

25 min

Module 2

Function Calling and Retrieval

Building knowledge-aware agents: function calling for arXiv search plus embeddings for ranking relevant documents.

Function Calling for Knowledge Retrieval

Build an agent with two functions: arXiv article search using embeddings, and PDF summarization. The agent decides which function to call based on the user's query.

25 min

Module 3

Embeddings and Search

Hands-on embeddings: indexing Wikipedia articles, chunking into sections, and building semantic search.

Embedding Wikipedia Articles for Semantic Search

Download Wikipedia articles about the 2022 Olympics, split into chunks, vectorize with text-embedding-ada-002, and build a semantic search engine over the knowledge base.

25 min

Module 4

Prompting and Optimization

Meta-prompting: automatic prompt optimization using a more powerful model to improve output quality.

Meta-Prompting: Automated Prompt Optimization

Use o1 to improve simple prompts: compare summarization quality with a basic and meta-optimized prompt on the BBC News dataset, processed in parallel with ThreadPoolExecutor.

25 min

Module 5

Vision and Multimodality

Processing video through GPT vision: frame extraction, scene analysis, and voiceover generation via the TTS API.

GPT Vision for Video Understanding

Extract frames from a video using OpenCV, send them to GPT-4.1-mini as base64 images, get a scene description, and generate a David Attenborough-style voiceover via the TTS API.

25 min

Module 6

Reliability and Quality

Building hallucination guardrails: eval dataset construction, judge model creation, and precision/recall measurement.

Hallucination Guardrails for Support Agents

Build an output guardrail for customer support: generate an eval set with GPT-4o, create a judge model to check whether responses follow policies, and measure precision and recall.

25 min

Module 7

Data Extraction and Transformation

GPT-4o as an OCR replacement: extracting structured data from multilingual PDF invoices and loading into a database via an ELT pipeline.

Data Extraction from PDFs: GPT as OCR

Build an ELT pipeline: PDF invoices → base64 → GPT-4o extracts raw JSON → schema transformation → SQLite. GPT-4o handles multilingual text and complex layouts better than traditional OCR.

25 min

Module 8

Fine-Tuning

Fine-tuning gpt-3.5-turbo for function calling: baseline evaluation, synthetic training data generation, and launching a fine-tuning job.

Fine-Tuning for Accurate Function Calling

Fine-tune gpt-3.5-turbo on synthetic data generated by gpt-4o for drone control: assess baseline accuracy, generate golden training examples, run a fine-tuning job, and compare results.

25 min