app.py

import nest_asyncio
import streamlit as st
import os
from flashrank import Ranker, RerankRequest
from qdrant_client import QdrantClient
from llama_index.llms.groq import Groq
from llama_index.core.retrievers import VectorIndexRetriever
from llama_index.retrievers.bm25 import BM25Retriever
from llama_index.core.retrievers import QueryFusionRetriever
from llama_index.core import VectorStoreIndex
from llama_index.vector_stores.qdrant import QdrantVectorStore
from llama_index.core import Settings
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
import PyPDF2
nest_asyncio.apply()
os.environ["HF_TOKEN"] = st.secrets["HF_TOKEN"]
groq_token = st.secrets["groq_token"]
st.set_page_config(
    layout="wide"
)

# default llamaindex llm and embendding model selection
@st.cache_resource(show_spinner=False)
def llamaindex_default():
    Settings.llm = Groq(model="llama-3.1-8b-instant", api_key=groq_token)
    Settings.embed_model = HuggingFaceEmbedding(
        model_name="law-ai/InLegalBERT", trust_remote_code=True
    )
llamaindex_default()

# set up qdrant client
@st.cache_resource(show_spinner=False)
def load_index():
    qdrant_client = QdrantClient(
        path="."
    )
    vector_store = QdrantVectorStore(
        client=qdrant_client, collection_name="legal_v1"
    )
    return VectorStoreIndex.from_vector_store(vector_store=vector_store)

index = load_index()

# reranker selection in the sidebar
with st.sidebar:
    selected_reranker = st.selectbox(
        "Select a reranker",
        ("default", "ms-marco-MiniLM-L-12-v2", "rank-T5-flan")
    )

    if selected_reranker == "default":
        ranker = Ranker()
    else:
        ranker = Ranker(model_name=selected_reranker, cache_dir=".")

    # Calculate individual weightages with sidebar slider 
    dense_weightage = st.slider("Dense Weightage", min_value=0.0, max_value=1.0, value=0.5, step=0.1)
    sparse_weightage = 1 - dense_weightage
    st.write("dense weight: ",dense_weightage)
    st.write("sparse weight: ",sparse_weightage)
    num_k = st.number_input(
        "Enter k",
        min_value=1,
        max_value=10,
        value=10
    )

@st.cache_resource(show_spinner=False)
def load_retriver():
    dense_retriever = VectorIndexRetriever(
        index=index,
        similarity_top_k=num_k
    )
    sparse_retriever = BM25Retriever.from_persist_dir("./sparse_retriever")
    sparse_retriever.similarity_top_k = num_k

    retriever = QueryFusionRetriever(
    [
        dense_retriever,
        sparse_retriever,
    ],
    num_queries=1,
    use_async=False,
    retriever_weights=[dense_weightage, sparse_weightage],
    similarity_top_k=num_k,
    mode="relative_score",
    verbose=True,
    )
    return retriever

retriever = load_retriver()

def extract_pdf_content(pdf_file_path):
    with open(pdf_file_path, 'rb') as pdf_file:
        pdf_reader = PyPDF2.PdfReader(pdf_file)
        text = ""
        for page_num in range(len(pdf_reader.pages)):
            page = pdf_reader.pages[page_num]
            text += page.extract_text()
        return text

#prompt template for summarization
template = """
Please summarize the following legal document and provide the summary in the specified format. The output should directly follow the format without any introductory text.
**Document:**
{document_content}

**Format:**

**Case:** [Case Number]

**Petitioner:** [Petitioner's Name]

**Respondent:** [Respondent's Name]

**Judge:** [Judge's Name]

**Order Date:** [Order Date]

**Summary:**
- **Background:** [Brief description of the case background]
- **Allegations:** [Summary of the allegations made in the case]
- **Investigation:** [Key findings from the investigation]
- **Court's Decision:** [Summary of the court's decision and any conditions imposed]
"""

st.title("Legal Documents Hybrid+Reranker Search")

query = st.text_input("Search through documents by keyword", value="")

search_btn = st.button("Search")

if search_btn and query:
    nodes = retriever.retrieve(query)
    passages = []
    for node in nodes:
        passage = {
            "id": node.node_id,
            "text": node.text,
            "meta": node.metadata
        }
        passages.append(passage)
    rerankrequest = RerankRequest(query=query, passages=passages)
    results = ranker.rerank(rerankrequest)  

    for node in results:
        st.write("File Name: ", node["meta"].get("file_name"))
        st.write("reranking score: ", node["score"])
        st.write("node id", node["id"])
        with st.expander("See Summary"):
            text = extract_pdf_content("./documents/" + node["meta"].get("file_name"))
            formatted_template = template.format(document_content=text)
            summary = Settings.llm.complete(formatted_template)
            st.markdown(summary)
        st.write("---")