06 — Labs

Hands-on Jupyter notebooks that turn Q&A knowledge into demonstrable, runnable skill. Each lab takes 30–60 minutes and includes a small bundled corpus.

Contents

Notebook	Run it	What You Build	Key Skills
01_naive_rag.ipynb		End-to-end naive RAG: chunk → embed → FAISS → Claude	Indexing pipeline, cosine retrieval, generation
02_hybrid_rag.ipynb		BM25 + dense retrieval merged with RRF	rank_bm25, FAISS, RRF formula, comparison vs. each alone
03_reranker_pipeline.ipynb		Two-stage: bi-encoder → cross-encoder reranker	CrossEncoder, latency trade-off, rank change analysis
04_ragas_evaluation.ipynb		Full eval harness: RAGAS + custom judges + regression detection	Golden datasets, RAGAS 4 metrics, LLM-as-judge, Recall@k
05_agentic_rag.ipynb		ReAct & Plan-and-Execute agentic retrieval loops with tool use, stopping criteria, and prompt-injection guardrails	Provider-agnostic function-calling loops, multi-hop query decomposition, agent evaluation, injection defense

Each notebook's first code cell installs its own dependencies and (when opened in Colab) fetches ai_client.py and prompts for an API key — no local setup required to try a lab. For the full local experience (all 5 labs, one shared environment), follow the steps below.

Prerequisites

1. Create a virtual environment

Windows (PowerShell):

py -3 -m venv .venv
.\.venv\Scripts\Activate.ps1

macOS/Linux:

python3 -m venv .venv
source .venv/bin/activate

2. Install dependencies

pip install --upgrade pip
pip install -r requirements.txt

3. Register the environment as a Jupyter kernel

python -m ipykernel install --user --name rag-labs --display-name "Python (rag-labs)"

Select the Python (rag-labs) kernel when opening a notebook (in VS Code: top-right kernel picker; in Jupyter: Kernel → Change Kernel).

4. Configure your LLM provider

Copy .env.example to .env and set AI_PROVIDER, AI_MODEL, and AI_API_KEY (Labs 02-05 read these via ai_client.py, which supports claude, gemini, openai, azure, and ollama). Lab 01 additionally needs HF_TOKEN (a HuggingFace Hub token), independent of AI_PROVIDER.

Copy-Item .env.example .env

cp .env.example .env

Then edit .env with your provider and key.

Lab Progression

Lab 01 (Naive RAG)
  └── Learn: indexing + retrieval + generation baseline

Lab 02 (Hybrid RAG)
  └── Add: BM25 + RRF on top of Lab 01 dense index
  └── Why: exact-match queries that dense alone misses

Lab 03 (Reranker)
  └── Add: cross-encoder reranking on top of Lab 02 candidates
  └── Why: rank precision when fast retrieval makes ordering mistakes

Lab 04 (Evaluation)
  └── Measure: RAGAS + Recall@k on Labs 01–03
  └── Why: without measurement, improvement is guesswork

Lab 05 (Agentic RAG)
  └── Add: multi-step ReAct / Plan-and-Execute tool-use loops on top of Lab 01–03 retrieval
  └── Why: single-shot pipelines can't handle compound, multi-hop queries requiring reasoning + tool use

Estimated Costs

Labs 02-05 use whichever model AI_MODEL is set to (see .env); costs above assume a small/cheap model like gemini-2.5-flash-lite. Lab 01 uses HuggingFace's Mistral-7B (or local Ollama), independent of AI_PROVIDER.