OCR-MODEL

A complete pipeline for building a Retrieval-Augmented Generation (RAG) system using PDFs as input. This project uses marker-pdf for accurate text extraction with layout awareness, numpy for data handling, and LLMs for intelligent question-answering.

---

Features

• Merge and save multiple PDFs into a single file.
• Load and extract text using marker-pdf (with layout preservation).
• Support for both single and batch PDF extraction.
• Convert extracted .md files into text chunks using recursive chunking.
• Generate sentence embeddings from chunks and save as .npy.
• Perform contextual search and Q&A using RAG (LLM-based).

---

Project Structure

 pdf-rag
│
├── pdfs/                    # Input PDF files
├── merged.pdf               # Merged PDF file (optional)
├── extracted_markdowns/    # Marker-pdf generated .md files
├── chunks/                 # Chunked text files
├── embeddings/             # Saved numpy arrays (.npy) for chunks
├── rag.py                  # Main script for querying with RAG
├── requirements.txt
└── README.md               # This file

---

Dependencies

• marker-pdf
• numpy
• sentence-transformers
• langchain or llama-index (based on LLM setup)
• PyPDF2 (for merging)
• markdown, tiktoken (for chunking if needed)
• Your preferred LLM backend (e.g., OpenAI, LLama, Gemini, etc.)

---

1. Save PDFs into one file

Use a Python script or tool like PyPDF2 to merge:

from PyPDF2 import PdfMerger

merger = PdfMerger()
for file in os.listdir("pdfs"):
    merger.append(f"pdfs/{file}")
merger.write("merged.pdf")
merger.close()

---

2. Extract text using marker-pdf

Single PDF:

!marker_single merged.pdf

Multiple PDFs:

!marker pdfs/

Markdown files will be saved in the current directory or specified output folder.

---

3. Chunk the .md files

Use a recursive chunking logic (e.g., based on headings, sentences, or token count) and save the output chunks.

---

4. Generate embeddings

Convert each chunk into a dense vector using a sentence-transformer model:

from sentence_transformers import SentenceTransformer
import numpy as np
import os

model = SentenceTransformer("all-MiniLM-L6-v2")

chunks_dir = 'chunks/'
embeddings = []

for file in os.listdir(chunks_dir):
    with open(os.path.join(chunks_dir, file), 'r', encoding='utf-8') as f:
        text = f.read()
        emb = model.encode(text)
        embeddings.append(emb)

np.save("embeddings/embeddings.npy", embeddings)

---

5. Build rag.py

script :import os import requests import numpy as np from scipy.spatial import distance from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

txtdir = 'chunk-texts as npy' embdir = 'Embeddings as npy'

combined_chunks = [] combined_embs = []

for f in os.listdir(txtdir): embfile = os.path.join(embdir, f) txtfile = os.path.join(txtdir, f)

if os.path.exists(embfile) and os.path.exists(txtfile):
    try:
        embs = np.load(embfile)
        chunks = np.load(txtfile, allow_pickle=True)
        if len(embs) != len(chunks):
            continue
        combined_chunks.extend(chunks)
        combined_embs.extend(embs)
    except Exception:
        continue
else:
    continue

Query = 'What is residual prophet inequality?' try: query_emb = model.encode(Query, convert_to_numpy=True) except Exception: query_emb = None

if query_emb is not None and combined_embs: try: combined_embs_array = np.array(combined_embs) distances = distance.cdist([query_emb], combined_embs_array, 'cosine')[0] top_3_indices = np.argsort(distances)[:3]

    llm_prompt = f'Answer the user query based on following context. User query = {Query}\nContext:'
    counter = 1
    for idx in top_3_indices:
        ctx = combined_chunks[idx]
        llm_prompt += f'\nContext # {counter} - {ctx}'
        counter += 1

    print(llm_prompt)
except Exception:
    pass

• Loads embeddings
• Accepts user query
• Finds the most relevant chunk(s)
• Sends combined context + query to an LLM
• Returns an intelligent answer

---

Sample Usage/ output

python rag.py
# Query: "What is residual prophet inequality?"
# Answer: [ The prophet inequality problem, as introduced by [Krengel and Sucheston](#page-29-0) [\(1977\)](#page-29-0), was resolved by using a dynamic program that gave a tight approximation ratio of 1/2. [Samuel-Cahn](#page-30-0) [\(1984\)](#page-30-0) later proved that a single-threshold strategy yields the same guarantee; this also showed that the order in which the variables are observed is immaterial.]

---