The Stack We’re Building On

Before writing a single line of application code it is worth spelling out exactly which packages this book uses, and which packages I have deliberately left out. The goal is that after reading this chapter you can look at any pyproject.toml in the source-code/ directory and know why every dependency is there.

Across every chapter in this book, the total set of Python packages you will install is small. Here it is, grouped by purpose:

Core frameworks

• langchain: the LangChain 1.0 core library.
• langchain-core: the primitives (Runnable, BaseMessage, PromptTemplate, etc.) that langchain and langgraph share.
• langgraph: the LangGraph 1.0 stateful-agent framework.
• langgraph-checkpoint-sqlite: the SqliteSaver checkpointer backend that makes a LangGraph agent restart-safe (Chapter “Durable, Restart-Safe Agents”).
• llama-index-core: the LlamaIndex 0.14+ core library.

Notice langchain-community is not on this list. Earlier editions leaned on it for SQL toolkits, retrievers, and cross-encoders; the version that shipped alongside LangChain 1.0 dropped several of those (SQLDatabaseToolkit, ContextualCompressionRetriever/CrossEncoderReranker, EnsembleRetriever, BM25Retriever, HuggingFaceCrossEncoder all lost their import paths at one point or another). Rather than pull in a large legacy package for pieces that keep moving, the chapters that used to need it (Chapters “RAG Patterns with LangChain” and “Natural-Language SQLite”) reimplement the specific piece directly on langchain_core.retrievers.BaseRetriever or raw SQLAlchemy, usually 20-40 lines. You will see why, with the actual code, in those chapters.

LLM providers (pick as needed per chapter)

• langchain-ollama and llama-index-llms-ollama: the default LLM provider for the book. Local models via Ollama.
• langchain-openai and llama-index-llms-openai: for readers who want to use OpenAI or an OpenAI-compatible endpoint such as Fireworks.ai. Used directly in the “swap the model” examples in Chapters “LangChain 1.0 in One Hour” and “LlamaIndex”.

Embeddings and vector stores

• langchain-huggingface and llama-index-embeddings-huggingface: local embedding models (BGE, nomic-embed).
• langchain-chroma: Chroma vector store integration, wrapping chromadb; a local vector store that persists to disk in a single directory.
• sentence-transformers: the reranker examples, and the raw pipeline behind the local-transformer LlamaIndex LLM example.
• rank-bm25 and llama-index-retrievers-bm25: sparse (keyword) retrieval for the hybrid RAG patterns, dense-plus-sparse in each framework.

Data-source and tool packages

• ddgs: the free web search backend used throughout.
• SPARQLWrapper: for the DBpedia and Wikidata chapter.
• trafilatura: for pulling clean text out of arbitrary web pages.
• pydantic: for structured-output schemas.
• pydrive2: Google Drive API access (the maintained fork of the abandoned pydrive).
• sqlalchemy: direct database access for the SQL agent chapter, which hand-rolls its tools on top of it instead of a toolkit.
• langchain-text-splitters: chunking text before embedding.

Local Hugging Face models and deployment

• transformers and torch: running a Hugging Face model directly through a local transformers pipeline, both standalone and wrapped as a custom LlamaIndex LLM.
• fastapi, uvicorn, and httpx: serving a LlamaIndex Workflow over plain HTTP, no managed deployment platform involved.

That list is exhaustive for the numbered chapters: every package above is a real dependency of at least one source-code/<chapter>/pyproject.toml, and every package in every chapter’s pyproject.toml is above. (A handful of older, standalone example directories predate this edition’s chapter structure and are not part of the book; they are not represented here.)

Every chapter’s source-code/<chapter>/pyproject.toml pins a version range, not an exact version such as langchain>=1.0,<2, pydantic>=2.7,<3, and so on, so a uv sync next year still gets compatible bugfixes. The part that actually pins exact versions is the uv.lock file next to it, generated the day the chapter was written and checked into the repo; uv sync reads the lock file by default, so what you install matches what was tested unless you deliberately uv lock --upgrade.

These are the packages that a search for “LangChain” or “LlamaIndex” will turn up first, and that we do not use anywhere in this book. It is worth naming them explicitly so you know why they are absent.

• langsmith: the client for LangSmith observability, evaluation, and prompt-hub features. LangSmith is a paid SaaS product ($39/seat/month for the Plus tier, plus per-trace usage). Everything the book does for observability we do with local tools like set_debug(True), callbacks, and (optionally) a self-hosted OpenInference collector.
• langgraph-sdk (cloud client): the client for the managed LangGraph deployment platform. We use langgraph itself, which is MIT-licensed and runs anywhere Python runs.
• langserve (in its managed form): where we serve LangGraph or LlamaIndex apps, we use plain FastAPI plus the framework’s own primitives.
• llama-cloud-services and llama-parse: the SDKs for LlamaCloud and LlamaParse. Both are credits-based commercial services. Where we need document parsing we use pypdf, unstructured, or Python’s own text handling, plus the OSS ingestion pipeline built into llama-index-core.
• llama-index-agent-openai as the primary agent story: LlamaIndex now recommends the Workflows API for new agent development, and that is what Part II uses.

If you already pay for one or more of these platforms and want to use them alongside the code in this book, nothing here will stop you. But nothing here will require you to.

Every chapter’s example directory looks like this:

To run any example, cd into its directory and do:

uv sync creates a .venv in the chapter directory and installs from uv.lock if one is present (every chapter ships one) or resolves fresh from pyproject.toml otherwise. uv run executes a script inside that venv. You never need to source .venv/bin/activate manually, and you never need to think about which Python interpreter is active.

If you prefer plain pip, each chapter’s pyproject.toml is a normal PEP 621 file and works with pip install -e . in a manually created virtual environment. The book uses uv for its snippets because it is faster and it avoids the “which environment am I in” class of bugs that used to eat an hour of every new reader’s weekend.

If you do these three things once, every chapter in the book will work:

1. Install uv (brew install uv on macOS, or curl -LsSf https://astral.sh/uv/install.sh | sh on Linux).
2. Install Ollama from ollama.com and pull one general-purpose model that supports tool calling. As of mid-2026 the models I use most often for the book examples are qwen3.5:4b for tool-calling work, llama3.2:3b when I want something small and fast, and gemma4:12b-it-qat when I have sufficient headroom on a 16 GB or 32G machine.
3. Optionally, put a .env file in your home directory with any hosted-model API keys you want to use:

That is the whole setup. You are ready for Part I.