Module 4 – Retrieval-Augmented Generation (RAG)

🚀 Quick Summary

Built a RAG pipeline that turns my resume/portfolio into a searchable knowledge base and answers questions with real citations.
Implements chunking → embeddings → vector index → retrieval → LLM generation via LangChain.
Runs with OpenAI or a local vLLM endpoint; supports FAISS or Chroma.
Demonstrates NLP preprocessing, vector search, LLM integration, and evaluation of retrieval quality.

---

📖 Project Description

Modern LLMs are powerful but can be out-of-date or hallucinate. This project makes answers grounded in my actual documents by combining LLMs with a real-time retrieval layer.
It converts resume and portfolio files into embeddings, indexes them in a vector DB, retrieves the most relevant chunks, and augments LLM responses with those snippets (and optional citations). The repo also includes a Python class demo that walks through the full pipeline and simple evaluation.

---

🎯 Objectives

• Ingestion & Chunking: Load resume (PDF/DOCX) + extras; split via RecursiveCharacterTextSplitter (~500 tokens, ~10% overlap).
• Embeddings & Indexing: Create embeddings (OpenAI or HF) and store in FAISS or Chroma.
• Retrieval & QA: Build LangChain RetrievalQA to fetch top-k chunks and generate answers with references.
• Evaluation: Probe with good/bad/adversarial queries; compute simple metrics (e.g., context recall, accuracy proxy) and discuss nDCG/attribution.
• (Optional) Local Serving: Connect to vLLM (e.g., Zephyr-7B) via OpenAI-compatible endpoint for GPU inference.

---

🛠️ Tech Stack

• Language: Python
• Frameworks/Libraries: LangChain, FAISS or Chroma, OpenAI API / vLLM (OpenAI-compatible), Hugging Face
• Key Modules:
  • Embeddings: OpenAIEmbeddings or HuggingFaceEmbeddings
  • Retrieval: FAISS / Chroma vector stores
  • RAG Chain: RetrievalQA (LangChain)
  • Utilities: dotenv, pypdf, TextLoader, RecursiveCharacterTextSplitter
• (GPU Option): vLLM server (OpenAI-compatible), sentence-transformers for local embeddings

---

🔥 Architecture / Workflow Diagram

flowchart LR A[Docs]-->B[Chunk] B-->C[Embeddings] C-->D[FAISS/Chroma] E[Query]-->F[Retriever(k)] D-->F F-->G[LLM Generate + Cite]

---

📂 Deliverables

• class_4.py / notebooks/ — end-to-end RAG pipeline (load → chunk → embed → index → retrieve → answer → evaluate)
• resume.pdf + portfolio_notes.txt — sample inputs for the Resume AI use case
• vector_index/ — persisted FAISS/Chroma index (if saved)
• examples/qa_samples.jsonl — example questions and answers
• reports/eval_stats.md — notes on retrieval quality (e.g., accuracy proxy, recall, improvement ideas)

---

🔥 How to Run / Quick Start

pip install -r requirements.txt python rag_run.py --resume resume.pdf --notes portfolio.txt --k 3

---

🌟 Highlights

• End-to-end RAG with clear, modular steps (easy to swap embeddings/DBs/LLMs).
• Grounded answers with retrieved snippets (reduces hallucinations).
• Local or cloud: works with OpenAI or local vLLM for cost/privacy.
• Evaluation-aware: provides a baseline harness and discusses metrics (recall, nDCG, attribution).
• Career-ready demo: “My Resume AI” answers recruiter-style questions from real documents.

---

🧭 Typical Workflow

1. Load Docs: PyPDFLoader("./resume.pdf") + TextLoader("./portfolio_notes.txt")
2. Split: RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
3. Embed & Index: FAISS.from_documents(chunks, OpenAIEmbeddings(...)) (or Chroma.from_documents)
4. RAG Chain:

   agent = RetrievalQA.from_chain_type(
     llm=OpenAI(temperature=0),  # or ChatOpenAI(openai_api_base="http://localhost:8000/v1", ...)

---

🌟 Skills Demonstrated

RAG design；vector DB；prompting & evaluation；LLM deployment patterns。

---

 chain_type="stuff",
 retriever=vectorstore.as_retriever(search_kwargs={"k": 3})

)