A global object instead of explicit passing
Every LangChain lesson in this repo builds a model object and passes it explicitly wherever it's needed: ChatGoogleGenerativeAI(...), then model.invoke(...), or hands it into create_agent(model=...). LlamaIndex does this differently by default: it has one global object, Settings, that most of the framework reads from automatically.
Set Settings.llm and Settings.embed_model once, and every Index, QueryEngine, and Agent you build for the rest of a script picks them up without you passing them in by hand each time. You can still override per call when you need to (later lessons do), but the global is the default, and it's the single biggest ergonomic difference between how these two frameworks feel to write.
| LangChain | LlamaIndex | |
|---|---|---|
| Model setup | model = ChatGoogleGenerativeAI(...) | Settings.llm = GoogleGenAI(...) |
| Passed to each call? | Yes, explicitly | No, read from the global by default |
| Embeddings setup | GoogleGenerativeAIEmbeddings(...), passed to a vector store | Settings.embed_model = GoogleGenAIEmbedding(...) |
| Override per call | Always explicit anyway | Possible, but the exception, not the rule |
The code, piece by piece
from llama_index.llms.google_genai import GoogleGenAIfrom llama_index.embeddings.google_genai import GoogleGenAIEmbeddingGoogleGenAI is LlamaIndex's LLM wrapper for Gemini, the equivalent of LangChain's ChatGoogleGenerativeAI. GoogleGenAIEmbedding is the embeddings equivalent of GoogleGenerativeAIEmbeddings.
Settings.llm = GoogleGenAI(model="gemini-3.5-flash-lite", api_key=API_KEY)Settings.embed_model = GoogleGenAIEmbedding(model_name="models/gemini-embedding-001", api_key=API_KEY)Same model names used throughout this repo: gemini-3.5-flash-lite for the LLM, models/gemini-embedding-001 for embeddings. Set once, at module level or in main(), before building any Index.
response = Settings.llm.complete("In one short sentence, what is a vector embedding?").complete() is a direct, index-free call to the LLM, proving the connection works before Lesson 4 builds a real index on top of it.
embedding = Settings.embed_model.get_text_embedding("hello world").get_text_embedding() turns one string directly into its vector, the same operation an Index runs internally on every Node, exposed here so you can see the shape of it: a list of floats, one number per dimension.
Checkpoint
Settings: a global object holding the LLM and embedding model most of LlamaIndex reads from automatically, LlamaIndex's alternative to LangChain's pattern of passing model objects explicitly everywhere.- Set
Settings.llmandSettings.embed_modelonce, before building anyIndex. GoogleGenAIandGoogleGenAIEmbeddingare LlamaIndex's wrappers for Gemini, matching the model names already used elsewhere in this repo..complete()and.get_text_embedding()are direct, index-free calls useful for testing a connection or understanding what an Index does internally.
If anything here still feels unclear, ask before moving to Lesson 4.