Environment Setup

1. Python and Dependencies#

Requires Python ≥ 3.12 and uv.

cd llm-engineering-program
uv sync --locked --all-extras
source .venv/bin/activate

uv sync --locked installs only the module 1–2 base. Use --all-extras so the RAG, agent, harness, and production labs never depend on packages that happen to exist on your machine.

Alternative without uv:

python3 -m venv .venv && source .venv/bin/activate
pip install -e '.[rag,agents,ops,dev]'

2. API Keys#

Copy the template and fill in only the keys you will use in each module:

cp setup/.env.example .env

Variable	Where to get it	Used from
OPENAI_API_KEY	platform.openai.com	Module 1
ANTHROPIC_API_KEY	console.anthropic.com	Module 1
AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_REGION	AWS Console (IAM)	Module 1 (Bedrock)
COHERE_API_KEY	dashboard.cohere.com	Module 3 (rerank)
LANGSMITH_API_KEY	smith.langchain.com	Module 8 (--send only)

The labs load .env with python-dotenv. Never upload .env to the repo (it is already in .gitignore).

3. Estimated API Costs#

Cost depends on the current catalog and pricing. The August 2026 examples use gpt-5.6-luna and claude-haiku-4-5 for volume, reserving higher tiers for justified comparisons. Check the official price before each run and set a budget; modules 3–5 also include offline workflows or local models with Ollama.

4. Optional Local Services#

• Ollama (brew install ollama) — local models for deployment labs and for working without API costs.
• Docker Desktop — required in module 8 (containers) and for local Qdrant/pgvector in module 3.
• Node ≥ 20 — only for the Next.js frontend in module 3.

5. Verification#

python setup/check_env.py --profile all

It checks the Python version and dependencies for the selected profile. For security, it only reports whether .env exists: it never loads, inspects, or prints the status of individual credentials.

6. Instructor Compatibility#

The main environment sets the current OpenAI SDK 3.x. Instructor 1.15.4, the version current as of 21-08-2026, declares openai<3 and rich<15; installing it in the same environment would make the lock unresolvable. The structured outputs lab uses the native API by default and keeps the comparison with Instructor in an isolated process:

uv run --no-project \
  --with-requirements setup/requirements-instructor.txt \
  python modulo-02-prompt-engineering/labs/05_structured_outputs_instructor.py \
  --backend instructor

This does not change the model (OPENAI_MODEL=gpt-5.6-luna); it only isolates incompatible SDK versions. The requirements file makes the decision reproducible.

RAGAS 0.4.3 pulls the same dependency. Additionally, langchain-community<0.4 is pinned: RAGAS still imports chat_models.vertexai, a module removed in the 0.4.x branch. Its live mode runs in an isolated manner with lexical retrieval to avoid mixing a second embedding stack:

uv run --no-project \
  --with-requirements setup/requirements-ragas.txt \
  python modulo-03-rag/labs/06_evaluacion_ragas.py \
  --ragas --lexical --limit 5

The --offline mode of the lab belongs to the main environment and does not require RAGAS or API keys.