📚 Mini RAG System

A lightweight Retrieval-Augmented Generation (RAG) pipeline that ingests a PDF, chunks it intelligently, embeds it with Google's Gemini embedding model, stores/searches it in Pinecone, and now answers questions end-to-end using a tool-calling LangChain agent.

---

📁 Folder Structure

mini-rag-system/
├── venv/                  # Virtual environment (not tracked in git)
├── .env                   # API keys (Google, Pinecone) — not tracked in git
├── .gitignore
├── fullstack_guide.pdf     # Source PDF to be ingested
├── main.py                 # Core script: ingestion + RAG agent
└── README.md

✨ What it does

1. Loads a PDF (fullstack_guide.pdf) using PyPDFLoader
2. Cleans the extracted text
3. Splits it into overlapping chunks using RecursiveCharacterTextSplitter
4. Embeds each chunk with Google's gemini-embedding-001 model
5. Stores the embeddings in a Pinecone vector index
6. Searches the index semantically (not just keyword matching!)
7. Retrieves + Generates: a LangChain agent calls a custom getcontext tool to pull relevant chunks from Pinecone, then uses Gemini 2.5 Flash to generate a grounded answer

---

🛠 Tech Stack

---

📦 Setup

1. Clone & install dependencies

pip install langchain langchain-google-genai langchain-pinecone langchain-community langchain-text-splitters pinecone python-dotenv pypdf

2. Create a .env file

GOOGLE_API_KEY=your_google_api_key_here
PINECONE_API_KEY=your_pinecone_api_key_here

3. Create your Pinecone index

Make sure you have an index named rag-index1 with:

• Dimension: 3072 (matches gemini-embedding-001 output)
• Metric: cosine

4. Add your PDF

Drop your PDF file in the project root and update the filename in the script:

loader = PyPDFLoader("./fullstack_guide.pdf")

5. Run it

python main.py

---

🔍 How retrieval works (low-level)

Under the hood, similarity search against Pinecone looks like this:

vector_store.similarity_search(query="version control with git", k=1)

This returns the most relevant chunk(s) from your PDF based on meaning, not exact keyword matches.

---

🤖 RAG Agent (query → retrieve → generate)

Instead of calling similarity_search manually, the project now wraps retrieval in a tool and hands it to a LangChain agent powered by Gemini 2.5 Flash:

@tool
def getcontext(query: str):
    """Use this tool to get more information to fulfill the user's request."""
    result = vector_store.similarity_search(query=query, k=2)
    return str(result)

model = ChatGoogleGenerativeAI(model="gemini-2.5-flash")

agent = create_agent(
    model=model,
    tools=[getcontext],
    system_prompt=(
        "You must ALWAYS call the getcontext tool first to retrieve "
        "relevant information before answering any question. "
        "Base your answer only on the retrieved context."
    )
)

response = agent.invoke({
    "messages": [HumanMessage("What is Fullstack development?")]
})

Why a tool instead of a direct call?

• The agent decides when to retrieve, rather than retrieval being hardcoded into every query.
• The system_prompt enforces a tool-first workflow, so the model is grounded in your Pinecone data instead of answering from its own training knowledge (reduces hallucination).
• This sets up the project for multi-tool agents later (web search, calculators, etc.) without changing the core retrieval logic.

---

⚠️ Known Quirks (a.k.a. things I learned the hard way)

• index.delete(delete_all=True) throws a NotFoundException if the namespace is empty/new — wrap it in a try/except if running on a fresh index.
• Vector writes to Pinecone aren't instantly searchable — add a short time.sleep() after add_documents() if you're querying right away.
• Chunk boundaries aren't perfect with character-based splitting — a chunk might blend two topics together if they're short and close together. This is normal; retrieval + LLM generation downstream usually compensates for it.
• create_agent() expects system_prompt, not prompt — passing prompt= raises a TypeError.
• The agent only calls getcontext if it judges the question needs it. For very generic questions, it may answer from its own knowledge unless the system prompt explicitly forces tool-first behavior (as done above).