import streamlit as st import os import requests import pdfplumber import chromadb import re from langchain.document_loaders import PDFPlumberLoader from langchain_huggingface import HuggingFaceEmbeddings from langchain_experimental.text_splitter import SemanticChunker from langchain_chroma import Chroma from langchain.chains import LLMChain from langchain.prompts import PromptTemplate from langchain_groq import ChatGroq from prompts import rag_prompt, relevancy_prompt, relevant_context_picker_prompt, response_synth # ----------------- Streamlit UI Setup ----------------- st.set_page_config(page_title="Blah", layout="centered") st.title("Blah-1") # ----------------- API Keys ----------------- os.environ["GROQ_API_KEY"] = st.secrets.get("GROQ_API_KEY", "") # ----------------- ChromaDB Persistent Directory ----------------- CHROMA_DB_DIR = "/mnt/data/chroma_db" os.makedirs(CHROMA_DB_DIR, exist_ok=True) # ----------------- Initialize Session State ----------------- if "pdf_loaded" not in st.session_state: st.session_state.pdf_loaded = False if "chunked" not in st.session_state: st.session_state.chunked = False if "vector_created" not in st.session_state: st.session_state.vector_created = False if "processed_chunks" not in st.session_state: st.session_state.processed_chunks = None if "vector_store" not in st.session_state: st.session_state.vector_store = None # ----------------- Improved Metadata Extraction ----------------- def extract_metadata(pdf_path): """Extracts title, author, emails, and affiliations from PDF.""" with pdfplumber.open(pdf_path) as pdf: metadata = pdf.metadata or {} # Extract title title = metadata.get("Title", "").strip() if not title and pdf.pages: text = pdf.pages[0].extract_text() title_match = re.search(r"(?i)title[:\-]?\s*(.*)", text or "") title = title_match.group(1) if title_match else text.split("\n")[0] if text else "Untitled Document" # Extract author author = metadata.get("Author", "").strip() if not author and pdf.pages: author_match = re.search(r"(?i)by\s+([A-Za-z\s,]+)", pdf.pages[0].extract_text() or "") author = author_match.group(1).strip() if author_match else "Unknown Author" # Extract emails emails = re.findall(r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}", pdf.pages[0].extract_text() or "") email_str = ", ".join(emails) if emails else "No emails found" # Extract affiliations affiliations = re.findall(r"(?:Department|Faculty|Institute|University|College|School)\s+[\w\s]+", pdf.pages[0].extract_text() or "") affiliation_str = ", ".join(affiliations) if affiliations else "No affiliations found" return title, author, email_str, affiliation_str # ----------------- PDF Selection ----------------- pdf_source = st.radio("Choose a PDF source:", ["Upload a PDF file", "Enter a PDF URL"], index=0, horizontal=True) if pdf_source == "Upload a PDF file": uploaded_file = st.file_uploader("Upload your PDF file", type=["pdf"]) if uploaded_file: st.session_state.pdf_path = "/mnt/data/temp.pdf" with open(st.session_state.pdf_path, "wb") as f: f.write(uploaded_file.getbuffer()) st.session_state.pdf_loaded = False st.session_state.chunked = False st.session_state.vector_created = False # ----------------- Process PDF ----------------- if not st.session_state.pdf_loaded and "pdf_path" in st.session_state: with st.spinner("🔄 Processing document... Please wait."): loader = PDFPlumberLoader(st.session_state.pdf_path) docs = loader.load() st.json(docs[0].metadata) # Extract metadata title, author, email_str, affiliation_str = extract_metadata(st.session_state.pdf_path) # Display extracted metadata st.subheader("📄 Extracted Document Metadata") st.write(f"**Title:** {title}") st.write(f"**Author:** {author}") st.write(f"**Emails:** {email_str}") st.write(f"**Affiliations:** {affiliation_str}") # Embedding Model model_name = "nomic-ai/modernbert-embed-base" embedding_model = HuggingFaceEmbeddings(model_name=model_name, model_kwargs={"device": "cpu"}, encode_kwargs={'normalize_embeddings': False}) # Convert metadata into a retrievable chunk metadata_text = f"Title: {title}\nAuthor: {author}\nEmails: {email_str}\nAffiliations: {affiliation_str}" metadata_doc = {"page_content": metadata_text, "metadata": {"source": "metadata"}} # Prevent unnecessary re-chunking if not st.session_state.chunked: text_splitter = SemanticChunker(embedding_model) document_chunks = text_splitter.split_documents(docs) document_chunks.insert(0, metadata_doc) # Insert metadata as a retrievable document st.session_state.processed_chunks = document_chunks st.session_state.chunked = True st.session_state.pdf_loaded = True st.success("✅ Document processed and chunked successfully!") # ----------------- Setup Vector Store ----------------- if not st.session_state.vector_created and st.session_state.processed_chunks: with st.spinner("🔄 Initializing Vector Store..."): st.session_state.vector_store = Chroma( persist_directory=CHROMA_DB_DIR, # <-- Ensures persistence collection_name="deepseek_collection", collection_metadata={"hnsw:space": "cosine"}, embedding_function=embedding_model ) st.session_state.vector_store.add_documents(st.session_state.processed_chunks) st.session_state.vector_created = True st.success("✅ Vector store initialized successfully!") # ----------------- Query Input ----------------- query = st.text_input("🔍 Ask a question about the document:") if query: with st.spinner("🔄 Retrieving relevant context..."): retriever = st.session_state.vector_store.as_retriever(search_type="similarity", search_kwargs={"k": 5}) retrieved_docs = retriever.invoke(query) context = [d.page_content for d in retrieved_docs] st.success("✅ Context retrieved successfully!") # ----------------- Run Individual Chains Explicitly ----------------- context_relevancy_chain = LLMChain(llm=llm_judge, prompt=PromptTemplate(input_variables=["retriever_query", "context"], template=relevancy_prompt), output_key="relevancy_response") relevant_context_chain = LLMChain(llm=llm_judge, prompt=PromptTemplate(input_variables=["relevancy_response"], template=relevant_context_picker_prompt), output_key="context_number") relevant_contexts_chain = LLMChain(llm=llm_judge, prompt=PromptTemplate(input_variables=["context_number", "context"], template=response_synth), output_key="relevant_contexts") response_chain = LLMChain(llm=rag_llm, prompt=PromptTemplate(input_variables=["query", "context"], template=rag_prompt), output_key="final_response") response_crisis = context_relevancy_chain.invoke({"context": context, "retriever_query": query}) relevant_response = relevant_context_chain.invoke({"relevancy_response": response_crisis["relevancy_response"]}) contexts = relevant_contexts_chain.invoke({"context_number": relevant_response["context_number"], "context": context}) final_response = response_chain.invoke({"query": query, "context": contexts["relevant_contexts"]}) # ----------------- Display All Outputs ----------------- st.markdown("### Context Relevancy Evaluation") st.json(response_crisis["relevancy_response"]) st.markdown("### Picked Relevant Contexts") st.json(relevant_response["context_number"]) st.markdown("### Extracted Relevant Contexts") st.json(contexts["relevant_contexts"]) st.markdown("### RAG Final Response") st.write(final_response["final_response"]) st.subheader("context_relevancy_evaluation_chain Statement") st.json(final_response["relevancy_response"]) st.subheader("pick_relevant_context_chain Statement") st.json(final_response["context_number"]) st.subheader("relevant_contexts_chain Statement") st.json(final_response["relevant_contexts"]) st.subheader("RAG Response Statement") st.json(final_response["final_response"])