RAG Systems and Prompt Engineering — Practical Implementations

A collection of 11 end-to-end implementations covering RAG pipelines, semantic search, prompt engineering, and model evaluation using the Google Gemini API.

---

What this covers

• RAG Pipeline — Full retrieval-augmented generation using Gemini embeddings and cosine similarity
• RAG Triad Evaluation — Faithfulness, relevance, and context precision scoring to measure and reduce hallucinations
• Semantic vs Keyword Search — BM25 keyword retrieval compared against vector embedding search
• Prompt Engineering — Zero-shot, few-shot, and Chain-of-Thought comparisons with quantified output differences
• Clinical Document Parsing — Extracting structured data from medical PDF reports using PyMuPDF
• Synthetic Dataset Generation — Automated Q&A bank creation exported to Excel using openpyxl
• LLM Parameter Analysis — Temperature and top_p comparisons for stochastic vs deterministic output

---

RAG pipeline architecture

Unstructured Data (PDF/TXT)
        |
   Text Chunking
        |
 Gemini Embedding-001
        |
  In-Memory Vector Store
        |
   User Query --> Query Embedding --> Cosine Similarity Search
                                              |
                                     Top-K Context Retrieval
                                              |
                                   Prompt: Context + Query
                                              |
                                    Gemini 1.5 Pro / Flash
                                              |
                                    Final Answer + RAG Triad Score

---

Tech stack

Python · Google Generative AI (Gemini SDK) · PyMuPDF · NumPy · openpyxl

---

Setup

git clone https://github.com/Soum-Code/GenAI-Lab-Experiments.git
cd GenAI-Lab-Experiments
pip install -r requirements.txt

Replace the API key placeholder in each script:

API_KEY = "your_gemini_api_key_here"

Get a free Gemini API key at aistudio.google.com.

---

Implementations

#	Focus	What it builds
1	LLM Parameters	Temperature and top_p comparison across generation tasks
2	QA Metrics	Automated toxicity, bias, and fluency scoring
3	Lexical Search	BM25 keyword retrieval and its limitations
4	Semantic Search	Vector embeddings with cosine similarity retrieval
5	Document AI	Clinical PDF parsing and structured data extraction
6	Dataset Generation	Synthetic Q&A export to Excel
7	RAG Pipeline	End-to-end documentation retrieval and answering
8	RAG Triad	Faithfulness, relevance, and context precision evaluation
11	Prompt Strategy	Zero-shot vs few-shot performance comparison
12	Chain-of-Thought	Reasoning path elicitation for logic tasks
13	Fine-Tuning Prep	Supervised fine-tuning data preparation logic

---

License

MIT