How nebius-academy/llm-engineering-essentials works
Project Type and Purpose
• This is an educational “course‐as‐code” repository that teaches LLM engineering through hands-on Jupyter notebooks and supporting Python modules.
• The end-to-end “NPC Factory” platform is built over six topics, incrementally covering LLM APIs, agent workflows, RAG, self-hosted LLMs, optimization, and fine-tuning.
File Structure
• Top‐level files: LICENSE, README.md
• Three topic directories (topic1, topic2, topic3) each containing:
• Jupyter notebooks (lessons and solutions) for interactive learning
• Python scripts (.py) implementing demo code or utilities (e.g., dungeon_navigation.py, thought_analysis.py, deep_research.py, hnsw.py, potion_shop_utils.py)
• Data assets (fantasy_items_descriptions.json, potion_shop.db, northelm.txt)
• No dedicated “frontend” or “backend” folders—this is a library of code modules invoked by notebooks.
Extract Architectural Insights from the README
• The course roadmap builds a platform (“NPC Factory”) through discrete projects per topic.
• Core functional goals:
a. Interact with LLM APIs (topic1)
b. Orchestrate agents and reasoning workflows (topic2)
c. Implement RAG with vector stores & databases (topic3)
• Future topics (4–6) layer self-hosting, optimization, monitoring, fine-tuning—though code for them is not in v1.
Main Components
a. Presentation Layer
• Jupyter Notebook Interfaces (the .ipynb files in topic1–3)
b. Application Logic Layer
• Python modules in each topic (e.g., dungeon_navigation.py, thought_analysis.py, hnsw.py, potion_shop_utils.py, deep_research.py)
c. Data Layer
• Local vector store (implemented by hnsw.py + hnsw_visualization.py)
• SQLite DB (potion_shop.db)
• Static JSON assets (fantasy_items_descriptions.json)
• Text resources (northelm.txt)
d. External Services
• Hosted LLM APIs (e.g., OpenAI, Anthropic)
• (Future) Self-hosted LLM inference endpoints
Relationships and Data‐Flows
• Notebooks import and invoke Python modules.
• Modules send requests to LLM APIs, receive responses, and post-process results.
• RAG modules (hnsw.py + potion_shop_utils.py) index data from JSON/DB into vector store; query vector store during reasoning.
• Results flow back to notebooks for visualization or further chaining (e.g., dungeon navigation decisions).
Architectural Patterns
• Layered Architecture: Presentation (Notebooks) → Business (Python modules) → Data (DB/Vector Store) → External (LLM)
• Modular Design: Each topic folder is a self-contained module teaching a specific capability
• Pipeline/Workflow Pattern: Chaining LLM prompts, agent orchestration, RAG retrieval
• Incremental build: you may overlay Topic2 components onto Topic1’s base, then Topic3 onto that, to illustrate evolution.
• Keep labels concise but descriptive: e.g., “hnsw.py (Vector Indexer),” “potion_shop.db (Context DB).”