93+ production-ready reference implementations for AI engineering patterns: RAG, agents, multimodal, voice, and fine-tuning with beginner-to-advanced roadmap.
Before you start
Python 3.10+ and pip. Specific examples may need additional API keys. You can browse without any setup.
The repo is organized into top-level folders: RAG, Agents, Evals, MCP, Fine-tuning, and more. Open the folder that matches what you are building. Skim the README in that folder before opening any code. The README tells you which example is the simplest entry point.
Install that example's requirements and run it on your own data, not the demo data. A reference implementation running on your actual input is more useful than one running on a pre-packaged example.
Open the example's main file in Claude Code and ask Claude to walk through what each section does, then adapt the parts that do not match your stack. The goal is to own the code, not copy-paste it.
Wire the adapted code to your actual data source and run it against a real query. Most errors at this stage are missing environment variables or a library version mismatch. Read the error carefully before debugging.
The major pattern categories in the hub and when to reach for each.
| Command / Skill | What it does | When to use it |
|---|---|---|
RAG pipelines | Chunk → embed → retrieve → generate reference implementations for document Q&A | Adding question-answering over private documents or knowledge bases |
Agentic loops | LLM-driven planning and tool-use loops with self-correction | Tasks requiring multi-step reasoning, tool calls, or iterative refinement |
Structured outputs | Patterns for forcing JSON or typed responses via tool-call or response_format | Downstream code needs to parse the model's output reliably |
Eval harness | Automated testing of LLM outputs against ground-truth or rubrics | Shipping a model or prompt change and need to verify no regression |
MCP examplesadvanced | Model Context Protocol server reference implementations | Building a custom MCP server and need a working starting point |
Multi-agent patternsadvanced | Orchestrator → worker setups with shared context and message passing | Tasks too large for a single context window or that parallelize naturally |