🧠 Advanced RAG Cookbook

A practical collection of 10 Retrieval-Augmented Generation techniques — with working code, real benchmarks, and plain-English explanations.

RAG · Advanced RAG · RAG Techniques · LlamaIndex · Hybrid RAG · HyDE · Fusion RAG · Parent-Child RAG · Retrieval Augmented Generation

---

What is this?

This cookbook covers 10 RAG techniques — from the simplest naive baseline to multi-stage reranking pipelines. Each technique has:

• A README that explains what it does, why it works, and when to use it (no fluff)
• A standalone Python script you can run directly
• Real benchmark numbers from actual tests (not estimates)

Built with LlamaIndex + OpenAI. Tested against the same document set across all techniques so the comparisons are fair.

If you've ever Googled "which RAG technique should I use" and got a vague answer — this repo is the practical answer to that question.

---

Techniques

#	Technique	Accuracy	Avg Latency	Best For
01	Naive RAG	Medium	~4.7s	Baseline, simple docs, getting started
02	Unstructured RAG	Medium	~4.9s	Scanned PDFs, images, OCR-heavy files
03	Contextual Compression	High	~10s	Long docs, noisy datasets, high precision
04	Fusion RAG	High	~7.5s	Ambiguous queries, better recall
05	Hybrid RAG ⭐	High	~5.6s	Best all-around — keyword + semantic
06	HyDE RAG	High	~12.5s	Short/vague queries, Q&A over profiles
07	Parent-Child RAG	Very High	~8.3s	Long documents, context-heavy answers
08	RRR RAG	High	~5.3s	Conversational queries, chatbot interfaces
09	Sentence Compression	High	~6-7s	Noisy chunks, reducing context tokens
10	Rerank + Compress	Very High	~10-11s	Max precision, high-stakes QA

---

Quick Decision Guide

Default choice with no constraints → Hybrid RAG — handles keyword and semantic queries, reliable across document types.

Users type conversationally → RRR RAG — rewrites vague queries before retrieval.

Scanned PDFs or images → Unstructured RAG — OCR layer handles what text readers can't.

Short, vague questions → HyDE RAG — generates a hypothetical answer and retrieves against that.

Long documents, need context in answers → Parent-Child RAG — small chunks for precision, large chunks for context.

Accuracy is the only priority → Rerank + Compress — two-stage post-processing, cleanest context.

Just learning RAG → Naive RAG — understand the baseline before adding complexity.

---

Setup

1. Clone

git clone https://github.com/Jeevav62/Rag-techniques.git
cd Rag-techniques

2. Install dependencies

pip install -r requirements.txt

3. Add your OpenAI API key

cp .env.example .env
# Edit .env → add your OPENAI_API_KEY

4. (For OCR) Install Tesseract

Only needed for Unstructured RAG when working with scanned PDFs/images.

OS	Command
macOS	brew install tesseract
Ubuntu	sudo apt-get install tesseract-ocr
Windows	Download installer

---

Running a Technique

Each techniques/XX/ folder has a standalone script. Update SAMPLE_FILES and SAMPLE_QUERY at the top, then:

cd techniques/05_hybrid_rag
python hybrid_rag.py

Output includes: answer, latency, token count, and the exact chunks retrieved.

---

Interactive Demo (Gradio)

Try all 10 techniques side by side with your own documents:

python app.py

Opens at http://localhost:7860 — upload any PDF/doc/image, pick a technique, compare outputs.

---

Benchmark Results

All techniques tested on the same personal document QA dataset (bio/resume-style questions — "where does this person work?", "what city are they in?", etc.)

Technique	Accuracy	Avg Latency	Extra LLM Calls
Naive RAG	Medium	4.7s	0
Unstructured RAG	Medium	4.9s	0
Contextual Compression	High	10.0s	1 (reranker)
Fusion RAG	High	7.5s	1 (query expansion)
Hybrid RAG	High	5.6s	0
HyDE RAG	High	12.5s	1 (hypothesis gen)
Parent-Child RAG	Very High	8.3s	0
RRR RAG	High	5.3s	1 (query rewrite)
Sentence Compression	High	6-7s	0
Rerank + Compress	Very High	10-11s	1 (reranker)

---

Related

Before RAG comes chunking. If you want to understand how documents should be split before retrieval, check out the companion repo: Chunking Techniques — 6 strategies from fixed-size to agentic, with the same format as this cookbook.

---

Stack

Component	Library
RAG Framework	LlamaIndex 0.10+
LLM	OpenAI GPT-4o-mini
Embeddings	OpenAI text-embedding-3-large
Document Parsing	Unstructured
Sparse Retrieval	BM25 (rank-bm25)
UI	Gradio

---

Repo Structure

Rag-techniques/
├── app.py                          # Gradio demo — all 10 techniques
├── requirements.txt
├── .env.example
├── rag/                            # Package imported by app.py
│   ├── naive_rag.py
│   ├── unstructured_rag.py
│   ├── contextual_rag.py
│   ├── fusion_rag.py
│   ├── hybrid_rag.py
│   ├── hyde_rag.py
│   ├── parent_child_rag.py
│   ├── rrr_rag.py
│   ├── sentence_compression_rag.py
│   └── rerank_compress_rag.py
└── techniques/                     # Standalone scripts + docs per technique
    ├── 01_naive_rag/
    ├── 02_unstructured_rag/
    ├── 03_contextual_compression/
    ├── 04_fusion_rag/
    ├── 05_hybrid_rag/
    ├── 06_hyde_rag/
    ├── 07_parent_child_rag/
    ├── 08_rrr_rag/
    ├── 09_sentence_compression/
    └── 10_rerank_compress/

---

Contributing

Found a bug? Want to add a technique? PRs are welcome.

1. Fork the repo
2. Add your technique in techniques/XX_<name>/ with a README + script
3. Add the run_rag() function to rag/<name>.py and register it in rag/__init__.py and app.py
4. Open a PR

---

License

MIT — use it, modify it, share it.

---

If this helped you understand RAG better, consider giving it a ⭐