Enhance: Switched model to llama3.2, Improved RAG Capabilites for multi column text in document

Author: ToastCoder

dependencies.txt

@@ -4,6 +4,7 @@ ollama
 chromadb
 langchain
 langchain-community
-pypdf
 langchain-text-splitters
-langchain-ollama
+langchain-ollama
+pymupdf
+protobuf==3.20.3

main.py

@@ -1,17 +1,17 @@
 # Required Libs
-from langchain_community.document_loaders import PyPDFLoader
+from langchain_community.document_loaders import PyMuPDFLoader
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_community.vectorstores import Chroma
 from langchain_ollama import OllamaEmbeddings
 import ollama
 
 # Load and Chunk the Data
 print("[INFO] Reading HR Policy PDF...")
-loader = PyPDFLoader("rag_source/nasscom-hr-manual-2016.pdf") 
+loader = PyMuPDFLoader("rag_source/nasscom-hr-manual-2016.pdf") 
 pages = loader.load()
 
 # Break the dense PDF into small, 500-character chunks
-text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
+text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=150)
 chunks = text_splitter.split_documents(pages)
 
 # Embed and Store in Vector DB
@@ -21,7 +21,7 @@
 vectorstore = Chroma.from_documents(documents=chunks, embedding=embeddings)
 
 # Retriever to fetch the top 3 most relevant chunks
-retriever = vectorstore.as_retriever(search_kwargs={"k": 3})
+retriever = vectorstore.as_retriever(search_kwargs={"k": 2})
 
 print("\n\nHi there! (Press Ctrl+C to exit)")
 
@@ -36,16 +36,19 @@
 
         # Construct the Prompt with Attention
         # Custom RAG Prompt to only use resources from the rag source
-        rag_prompt = f"""You are a helpful HR assistant. Answer the user's question using ONLY the context provided below. If the answer is not in the context, say you don't know.
+        rag_prompt = f"""You are a precise HR assistant. 
+        Read the provided Context carefully. You must answer the user's question using ONLY the facts found in the Context below. 
+        
+        If the Context does not contain the exact answer, simply reply: "I cannot find the answer to that in the current policy."
         
         Context: 
         {context_block}
         
         Question: {user_input}
         Answer:"""
 
-        # Generate the answer using TinyLlama
-        response = ollama.chat(model='tinyllama', messages=[
+        # Generate the answer using Llama 3.2
+        response = ollama.chat(model='llama3.2', messages=[
             {'role': 'user', 'content': rag_prompt}
         ])
 

run.sh

@@ -1,5 +1,3 @@
-#!/bin/bash
-
 # Variables
 PKG_MGR="pip"
 FILE="dependencies.txt"
@@ -73,9 +71,23 @@ else
     echo "[OK] Ollama engine is already installed."
 fi
 
+# Check for PyMuPDF
+if ! python3 -c "import pymupdf" &> /dev/null; then
+    echo "[WARNING] Python cannot find 'pymupdf' library. Forcing strict inline installation..."
+    python3 -m pip install pymupdf
+    
+    if [ $? -ne 0 ]; then
+        echo "[ERROR] Could not force-install Python pymupdf library."
+        exit 1
+    fi
+    echo "[OK] 'pymupdf' bound to current Python environment."
+else
+    echo "[OK] Python recognizes 'pymupdf'."
+fi
+
 # Required Models for RAG
-echo "[INFO] Verifying AI models (TinyLlama & Nomic)..."
-ollama pull tinyllama
+echo "[INFO] Verifying AI models (Llama3.2 & Nomic)..."
+ollama pull llama3.2
 ollama pull nomic-embed-text
 echo "[OK] All models are ready."