Basic RAG for PDF chat - Short & Crisp

Overview

So, RAG stands for Retrieval‑Augmented Generation. Basically, it’s a process of injecting external data into a prompt to get the desired response. External data could come from a database, the web, API calls, local files, etc.

What to do

Injection Phase

• Accept user pdf

• Make vector Embeddings

• Store it in a vector DB

Retrieval Phase

• Accept user query

• Make vector Embeddings

• Search in vector DB and Retrieve Relevant Chunks

Now, based on this relevant chunks + user query, we ask the LLM, and return its response back to user. That’s a basic typical RAG. You can choose any embeddings model, LLM’s and vector stores you prefer.

How to do

Loader

from langchain_community.document_loaders import PyPDFLoader

file_path = "./public/file_nane.pdf"
loader = PyPDFLoader(file_path)
docs = loader.load()

Chunking

from langchain_text_splitters import RecursiveCharacterTextSplitter

text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
split_docs = text_splitter.split_documents(loader.load())

Make vector Embeddings & Store it in a vector DB

from langchain_google_genai import GoogleGenerativeAIEmbeddings
from langchain_qdrant import QdrantVectorStore

embeddings = GoogleGenerativeAIEmbeddings(model="models/text-embedding-004")

qdrant = QdrantVectorStore.from_documents(
    documents=[], # empty docoment for first time. It will fill the embeddings here
    embedding=embeddings,
    collection_name="learning-genai",
    url=os.getenv("QDRANT_URL"),
    api_key=os.getenv("QDRANT_API_KEY"),
)

qdrant.add_documents(documents=split_docs)

retriver = QdrantVectorStore.from_existing_collection(
    embedding=embeddings,
    collection_name="learning-genai",
    url=os.getenv("QDRANT_URL"),
    api_key=os.getenv("QDRANT_API_KEY"),
)

Search in vector DB and Retrieve Relevant Chunks

from langchain_groq import ChatGroq
from langchain_core.prompts import ChatPromptTemplate

user_query = "take the user query as input"

relevent_chunks = retriver.similarity_search(query = user_query)
# for res in relevent_chunks:
#     print(f"{res.page_content} [{res.metadata}]")

llm = ChatGroq(
    model="llama-3.1-8b-instant",
    temperature=0,
)

prompt_template = ChatPromptTemplate([
    ("system", "you are a helpful AI asistant who responds based on the availabe context. {context}"),
    ("user", "give a short and crisp answer of this query {query}")
])

prompt = prompt_template.invoke({"context": relevent_chunks, "query": user_query})
response = llm.invoke(prompt)
print(response.content)