app.py

import os
import tempfile
import llama_index
import streamlit as st
from llama_index.core.workflow import (Event,StartEvent,StopEvent,Workflow,step,Context)
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
from llama_index.llms.groq import Groq
from llama_index.core import VectorStoreIndex
from llama_index.core.schema import QueryBundle
from llama_index.core.schema import NodeWithScore
from llama_index.core.response_synthesizers import CompactAndRefine
from llama_index.readers.web import SimpleWebPageReader
from llama_index.core import SimpleDirectoryReader
from llama_index.core.schema import NodeWithScore
from llama_index.core.postprocessor.llm_rerank import LLMRerank
from pyvis.network import Network
import chromadb

# Set API Key
os.environ["GROQ_API_KEY"] = st.secrets["GROQ_API_KEY"]

# Define Events
class RetrieverEvent(Event):
    nodes: list[NodeWithScore]

class RerankEvent(Event):
    nodes: list[NodeWithScore]

# Workflow Variations
class RAGWorkflowBasic(Workflow):
    @step
    def ingest(self, ctx: Context, ev: StartEvent) -> StopEvent:
        dir_path = ev.get("dirname")
        documents = SimpleDirectoryReader(dir_path).load_data()
        embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")
        index = VectorStoreIndex.from_documents(documents, embed_model=embed_model)
        return StopEvent(result=index)

    @step
    def retrieve(self, ctx: Context, ev: StartEvent) -> RetrieverEvent:
        index = ev.get("index")
        query = ev.get("query")
        retriever = index.as_retriever(similarity_top_k=2)
        nodes = retriever.retrieve(query)
        return RetrieverEvent(nodes=nodes)

    @step
    def synthesize(self, ctx: Context, ev: RetrieverEvent) -> StopEvent:
        llm = Groq(model="llama3-70b-8192")
        summarizer = CompactAndRefine(llm=llm)
        response = summarizer.synthesize(ctx.get("query"), nodes=ev.nodes)
        return StopEvent(result=response)

class RAGWorkflowRerank(RAGWorkflowBasic):
    @step
    def rerank(self, ctx: Context, ev: RetrieverEvent) -> RerankEvent:
        query_str = ctx.get("query")
        query_bundle = QueryBundle(query_str=query_str)
        llm_rerank = LLMRerank(top_n=2, llm=Groq(model="llama3-70b-8192"))
        reranked_nodes = llm_rerank.postprocess_nodes(ev.nodes, query_bundle)
        return RerankEvent(nodes=reranked_nodes)

# Function to Visualize Workflows
def visualize_workflow(workflow_class, filename):
    from llama_index.utils.workflow import draw_all_possible_flows
    from pyvis.network import Network
    
    # Manually ensure RerankEvent is included in the graph
    net = Network(directed=True, height="750px", width="100%")
    
    # Add StopEvent and RerankEvent nodes manually
    net.add_node("StopEvent", label="StopEvent", color="#FFA07A", shape="ellipse")
    net.add_node("RerankEvent", label="RerankEvent", color="#90EE90", shape="ellipse")
    
    # Visualize the entire flow
    try:
        draw_all_possible_flows(workflow_class, filename=filename)
    except AssertionError:
        st.error("Visualization error occurred, manually added RerankEvent node.")
        with open(filename, "r") as f:
            st.components.v1.html(f.read(), height=700)


# Streamlit UI
st.title("RAG Workflow Experimentation")
workflow_choice = st.selectbox("Choose Workflow:", ["Basic Workflow", "Workflow with Reranking"])

# Sidebar: Upload or Input PDFs
data_source = st.sidebar.radio("Data Source", ["Upload PDF", "Provide PDF URL"])
query = st.sidebar.text_input("Enter Query", "What is Fibromyalgia?")
temp_dir = tempfile.mkdtemp()

if data_source == "Upload PDF":
    uploaded_files = st.sidebar.file_uploader("Upload PDFs", accept_multiple_files=True, type=["pdf"])
    if uploaded_files:
        for file in uploaded_files:
            with open(os.path.join(temp_dir, file.name), "wb") as f:
                f.write(file.read())
        st.sidebar.success("PDFs Uploaded Successfully!")
elif data_source == "Provide PDF URL":
    pdf_url = st.sidebar.text_input("Enter PDF URL")
    if st.sidebar.button("Download PDF"):
        os.system(f"wget {pdf_url} -O {temp_dir}/downloaded.pdf")
        st.sidebar.success("PDF Downloaded Successfully!")

# Run Workflow
if st.button("Run Workflow"):
    if os.listdir(temp_dir):
        st.write("### Step 1: Ingesting Documents...")
        # Step 1: Ingest Documents
        embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")
        documents = SimpleDirectoryReader(temp_dir).load_data()
        index = VectorStoreIndex.from_documents(documents, embed_model=embed_model)

        st.write("### Step 2: Retrieving Documents...")
        # Step 2: Retrieve Documents
        retriever = index.as_retriever(similarity_top_k=2)
        nodes = retriever.retrieve(query)

        if workflow_choice == "Workflow with Reranking":
            st.write("### Step 3: Reranking Results...")
            # Step 3: Wrap query into QueryBundle and rerank
            query_bundle = QueryBundle(query_str=query)  # Wrap query into QueryBundle
            reranker = LLMRerank(top_n=2, llm=Groq(model="llama3-70b-8192"))
            nodes = reranker.postprocess_nodes(nodes, query_bundle)

        st.write("### Step 4: Synthesizing Response...")
        # Step 4: Synthesize Response
        summarizer = CompactAndRefine(llm=Groq(model="llama3-70b-8192"))
        response = summarizer.synthesize(query, nodes=nodes)
        st.markdown(f"### **Response:** {response}")

        # Workflow Visualization
        st.write("### Workflow Visualization")
        workflow_class = RAGWorkflowRerank if workflow_choice == "Workflow with Reranking" else RAGWorkflowBasic
        visualize_workflow(workflow_class, "workflow.html")
    else:
        st.error("No PDF files found. Upload or download a PDF first.")