# import os
# import gradio as gr
# from pinecone import Pinecone
# from sentence_transformers import SentenceTransformer
# from typing import List, Dict, Optional
# from langchain_google_genai import ChatGoogleGenerativeAI
# from langchain.chains.summarize import load_summarize_chain
# from langchain.prompts import PromptTemplate, ChatPromptTemplate
# from langchain.docstore.document import Document
# import time
# import asyncio
# import plotly.graph_objects as go
# from neo4j import GraphDatabase
# import networkx as nx
# from langchain_community.vectorstores import Neo4jVector
# from langchain.chains.summarize import load_summarize_chain
# from langchain.chains import LLMChain
# from langchain_google_genai import GoogleGenerativeAI, GoogleGenerativeAIEmbeddings
# class EnhancedLegalSearchSystem:
# def __init__(
# self,
# google_api_key: str,
# neo4j_url: str,
# neo4j_username: str,
# neo4j_password: str,
# embedding_model_name: str = "intfloat/e5-small-v2",
# device: str = "cpu"
# ):
# """Initialize the Enhanced Legal Search System"""
# # Initialize LLM
# self.llm = GoogleGenerativeAI(
# model="gemini-pro",
# google_api_key=google_api_key,
# temperature=0.1
# )
# # Initialize embeddings
# self.embeddings = GoogleGenerativeAIEmbeddings(
# model="models/embedding-001",
# google_api_key=google_api_key,
# task_type="retrieval_query"
# )
# # Initialize Neo4j connection
# self.neo4j_driver = GraphDatabase.driver(
# neo4j_url,
# auth=(neo4j_username, neo4j_password)
# )
# # Initialize vector store
# self.vector_store = Neo4jVector.from_existing_graph(
# embedding=self.embeddings,
# url=neo4j_url,
# username=neo4j_username,
# password=neo4j_password,
# node_label="Document",
# text_node_properties=["text"],
# embedding_node_property="embedding"
# )
# # Initialize additional embedding model for enhanced search
# self.local_embedding_model = SentenceTransformer(
# model_name_or_path=embedding_model_name,
# device=device
# )
# # Initialize prompts
# self.init_prompts()
# def __del__(self):
# """Cleanup Neo4j connection"""
# if hasattr(self, 'neo4j_driver'):
# self.neo4j_driver.close()
# def init_prompts(self):
# """Initialize enhanced prompts for legal analysis"""
# self.qa_prompt = ChatPromptTemplate.from_messages([
# ("system", """You are a legal expert assistant specializing in Indian law.
# Analyze the following legal context and provide a detailed, structured answer to the question.
# Include specific sections, rules, and precedents where applicable.
# Format your response with clear headings and bullet points for better readability.
# Context: {context}"""),
# ("human", "Question: {question}")
# ])
# self.map_prompt = PromptTemplate(
# template="""
# Analyze the following legal text segment:
# TEXT: "{text}"
# Instructions:
# 1. Extract and summarize the key legal points
# 2. Maintain all legal terminology exactly as written
# 3. Preserve section numbers and references
# 4. Keep all specific conditions and requirements
# 5. Include any mentioned time periods or deadlines
# KEY POINTS:
# """,
# input_variables=["text"] # Removed page_number as it's not used in the template
# )
# self.combine_prompt = PromptTemplate(
# template="""
# Question: {question}
# Using ONLY the information from the following legal document excerpts, provide a comprehensive answer:
# {text}
# Instructions:
# 1. Base your response EXCLUSIVELY on the provided document excerpts
# 2. If the documents don't contain enough information to fully answer the question, explicitly state what's missing
# 3. Use direct quotes when appropriate
# 4. Organize the response by relevant sections found in the documents
# 5. If there are conflicting statements across documents, highlight them
# ANALYSIS:
# """,
# input_variables=["text", "question"]
# )
# # Initialize summarize chain
# self.chain = load_summarize_chain(
# llm=self.llm,
# chain_type="map_reduce",
# map_prompt=self.map_prompt,
# combine_prompt=self.combine_prompt,
# verbose=True
# )
# def get_related_legal_entities(self, query: str) -> List[Dict]:
# """Retrieve related legal entities and their relationships"""
# # Corrected Cypher query to handle aggregation properly
# cypher_query = """
# // First, let's check if nodes exist and get their labels
# MATCH (d:Document)
# WHERE toLower(d.text) CONTAINS toLower($query)
# WITH d
# // Match all relationships from the document, collecting their types
# OPTIONAL MATCH (d)-[r]-(connected)
# WHERE NOT connected:Document // Avoid direct document-to-document relations
# WITH d,
# collect(DISTINCT type(r)) as relationTypes,
# collect(DISTINCT labels(connected)) as connectedLabels
# // Now use these to build our main query
# MATCH (d:Document)-[r1]-(e)
# WHERE toLower(d.text) CONTAINS toLower($query)
# AND NOT e:Document // Exclude direct document connections
# WITH d, r1, e
# // Get secondary connections, but be more specific about what we're looking for
# OPTIONAL MATCH (e)-[r2]-(related)
# WHERE (related:Entity OR related:Concept OR related:Section OR related:Case)
# AND related <> d // Prevent cycles back to original document
# WITH d, {
# source_id: id(d),
# source_text: d.text,
# document_type: COALESCE(d.type, "Unknown"),
# relationship_type: type(r1),
# entity: {
# id: id(e),
# type: CASE WHEN e:Entity THEN "Entity"
# WHEN e:Concept THEN "Concept"
# WHEN e:Section THEN "Section"
# WHEN e:Case THEN "Case"
# ELSE "Other" END,
# text: COALESCE(e.text, e.name, e.title, "Unnamed"),
# properties: properties(e)
# },
# related_entities: collect(DISTINCT {
# id: id(related),
# type: CASE WHEN related:Entity THEN "Entity"
# WHEN related:Concept THEN "Concept"
# WHEN related:Section THEN "Section"
# WHEN related:Case THEN "Case"
# ELSE "Other" END,
# relationship: type(r2),
# text: COALESCE(related.text, related.name, related.title, "Unnamed"),
# properties: properties(related)
# })
# } as result
# WHERE result.entity.text IS NOT NULL // Filter out any results with null entity text
# RETURN DISTINCT result
# ORDER BY result.source_id, result.entity.id
# LIMIT 25
# """
# try:
# with self.neo4j_driver.session() as session:
# # Execute the improved query
# result = session.run(cypher_query, {"query": query})
# entities = [record["result"] for record in result]
# # Log the results for debugging
# print(f"Found {len(entities)} related entities")
# if entities:
# for entity in entities:
# print(f"Entity: {entity['entity']['text']}")
# print(f"Source: {entity['source_text'][:100]}...")
# print(f"Related: {len(entity['related_entities'])} connections")
# return entities
# except Exception as e:
# print(f"Error in get_related_legal_entities: {str(e)}")
# return []
# async def process_legal_query(
# self,
# question: str,
# top_k: int = 5,
# context_window: int = 1
# ) -> Dict[str, any]:
# """Process a legal query using both graph and vector search capabilities"""
# try:
# # 1. Perform semantic search
# semantic_results = self.vector_store.similarity_search(
# question,
# k=top_k,
# search_type="hybrid"
# )
# # 2. Get related legal entities with the full question context
# related_entities = self.get_related_legal_entities(question)
# # Log the counts for debugging
# print(f"Found {len(semantic_results)} semantic results")
# print(f"Found {len(related_entities)} related entities")
# # 3. Expand context with related documents
# expanded_results = self.expand_context(
# semantic_results,
# context_window
# )
# # 4. Generate comprehensive answer
# documents = self._process_results(expanded_results, semantic_results)
# # 5. Prepare context for LLM
# context = self._prepare_context(documents, related_entities)
# # 6. Generate answer using LLM
# chain = LLMChain(llm=self.llm, prompt=self.qa_prompt)
# response = await chain.ainvoke({
# "context": context,
# "question": question
# })
# answer = response.get('text', '')
# # 7. Return structured response with explicit related concepts
# return {
# "status": "Success",
# "answer": answer,
# "documents": self._format_documents(documents),
# "related_concepts": related_entities, # This should now contain data
# "source_ids": sorted(list(set(doc.metadata.get('document_id', 'unknown') for doc in documents))),
# "context_info": {
# "direct_matches": len([d for d in documents if d.metadata.get('context_type') == "DIRECT MATCH"]),
# "context_chunks": len([d for d in documents if d.metadata.get('context_type') == "CONTEXT"])
# }
# }
# except Exception as e:
# print(f"Error in process_legal_query: {str(e)}") # Add error logging
# return {
# "status": f"Error: {str(e)}",
# "answer": "An error occurred while processing your query.",
# "documents": "",
# "related_concepts": [],
# "source_ids": [],
# "context_info": {}
# }
# def expand_context(
# self,
# initial_results: List[Document],
# context_window: int
# ) -> List[Document]:
# """Expand context around search results"""
# expanded_results = []
# seen_ids = set()
# for doc in initial_results:
# doc_id = doc.metadata.get('document_id', doc.page_content[:50])
# if doc_id not in seen_ids:
# # Query for related documents
# context_results = self.vector_store.similarity_search(
# doc.page_content,
# k=2 * context_window + 1,
# search_type="hybrid"
# )
# for result in context_results:
# result_id = result.metadata.get('document_id', result.page_content[:50])
# if result_id not in seen_ids:
# expanded_results.append(result)
# seen_ids.add(result_id)
# return expanded_results
# def _process_results(self, expanded_results: List[Document], initial_results: List[Document]) -> List[Document]:
# """Process and deduplicate search results"""
# seen_ids = set()
# documents = []
# for doc in expanded_results:
# doc_id = doc.metadata.get('document_id', doc.page_content[:50])
# if doc_id not in seen_ids:
# seen_ids.add(doc_id)
# is_direct_match = any(
# r.metadata.get('document_id', r.page_content[:50]) == doc_id
# for r in initial_results
# )
# doc.metadata['context_type'] = (
# "DIRECT MATCH" if is_direct_match else "CONTEXT"
# )
# documents.append(doc)
# return sorted(
# documents,
# key=lambda x: x.metadata.get('document_id', 'unknown')
# )
# def _prepare_context(
# self,
# documents: List[Document],
# related_entities: List[Dict]
# ) -> str:
# """Prepare context for LLM processing"""
# context = "\n\nLegal Documents:\n" + "\n".join([
# f"[Document ID: {doc.metadata.get('document_id', 'unknown')}] {doc.page_content}"
# for doc in documents
# ])
# if related_entities:
# context += "\n\nRelated Legal Concepts and Relationships:\n"
# for entity in related_entities:
# context += f"\n• {entity.get('entity', '')}"
# if entity.get('related_entities'):
# for related in entity['related_entities']:
# if related.get('entity'):
# context += f"\n - {related['type']}: {related['entity']}"
# return context
# def _format_documents(self, documents: List[Document]) -> str:
# """Format documents as markdown"""
# markdown = "### Retrieved Documents\n\n"
# for i, doc in enumerate(documents, 1):
# markdown += (
# f"**Document {i}** "
# f"(ID: {doc.metadata.get('document_id', 'unknown')}, "
# f"{doc.metadata.get('context_type', 'UNKNOWN')})\n"
# f"```\n{doc.page_content}\n```\n\n"
# )
# return markdown
# def generate_document_graph(
# self,
# query: str,
# top_k: int = 5,
# similarity_threshold: float = 0.5
# ) -> List[Dict]:
# """Generate graph data based on document similarity and relationships"""
# try:
# # 1. Get initial semantic search results
# semantic_results = self.vector_store.similarity_search(
# query,
# k=top_k,
# search_type="hybrid"
# )
# # 2. Get embeddings for all documents
# doc_texts = [doc.page_content for doc in semantic_results]
# doc_embeddings = self.local_embedding_model.encode(doc_texts)
# # 3. Create graph data structure
# graph_data = []
# seen_docs = set()
# # First, add all documents as nodes
# for i, doc in enumerate(semantic_results):
# doc_id = doc.metadata.get('document_id', f'doc_{i}')
# if doc_id not in seen_docs:
# seen_docs.add(doc_id)
# doc_type = doc.metadata.get('type', 'document')
# # Create node entry
# graph_data.append({
# 'source_id': doc_id,
# 'source_text': doc.page_content[:200], # Truncate for display
# 'document_type': doc_type,
# 'entity': {
# 'id': doc_id,
# 'type': 'Document',
# 'text': f"Document {i + 1}",
# 'properties': {
# 'similarity': 1.0,
# 'length': len(doc.page_content)
# }
# },
# 'related_entities': []
# })
# # Add relationships based on similarity
# from sklearn.metrics.pairwise import cosine_similarity
# similarity_matrix = cosine_similarity(doc_embeddings)
# # Create relationships between similar documents
# for i in range(len(semantic_results)):
# related = []
# for j in range(len(semantic_results)):
# if i != j and similarity_matrix[i][j] > similarity_threshold:
# doc_j = semantic_results[j]
# doc_j_id = doc_j.metadata.get('document_id', f'doc_{j}')
# related.append({
# 'id': doc_j_id,
# 'type': 'Document',
# 'relationship': 'similar_to',
# 'text': f"Document {j + 1}",
# 'properties': {
# 'similarity_score': float(similarity_matrix[i][j])
# }
# })
# # Add related documents to the graph data
# if related:
# graph_data[i]['related_entities'] = related
# return graph_data
# except Exception as e:
# print(f"Error generating document graph: {str(e)}")
# return []
# def create_graph_visualization(graph_data: List[Dict]) -> go.Figure:
# """Create an interactive graph visualization using Plotly"""
# if not graph_data:
# return go.Figure(layout=go.Layout(title='No documents found'))
# # Initialize graph
# G = nx.Graph()
# # Color mapping
# color_map = {
# 'Document': '#3B82F6', # blue
# 'Section': '#10B981', # green
# 'Reference': '#F59E0B' # yellow
# }
# # Node information storage
# node_colors = []
# node_texts = []
# node_hovers = [] # Full text for hover
# nodes_added = set()
# # Process nodes and edges
# for data in graph_data:
# source_id = data['source_id']
# source_text = data['source_text']
# # Add main document node
# if source_id not in nodes_added:
# G.add_node(source_id)
# node_colors.append(color_map['Document'])
# # Short text for display
# node_texts.append(f"Doc {len(nodes_added)+1}")
# # Full text for hover/click
# node_hovers.append(f"Document {len(nodes_added)+1}:
{source_text}")
# nodes_added.add(source_id)
# # Process related documents
# for related in data.get('related_entities', []):
# related_id = related['id']
# similarity = related['properties'].get('similarity_score', 0.0)
# if related_id not in nodes_added:
# G.add_node(related_id)
# node_colors.append(color_map['Document'])
# node_texts.append(f"Doc {len(nodes_added)+1}")
# node_hovers.append(f"Document {len(nodes_added)+1}:
{related['text']}")
# nodes_added.add(related_id)
# # Add edge with similarity weight
# G.add_edge(
# source_id,
# related_id,
# weight=similarity,
# relationship=f"Similarity: {similarity:.2f}"
# )
# # Create layout
# pos = nx.spring_layout(G, k=2.0, iterations=50)
# # Create edge trace
# edge_x = []
# edge_y = []
# edge_text = []
# for edge in G.edges(data=True):
# x0, y0 = pos[edge[0]]
# x1, y1 = pos[edge[1]]
# # Create curved line
# mid_x = (x0 + x1) / 2
# mid_y = (y0 + y1) / 2
# # Add some curvature
# mid_x += (y1 - y0) * 0.1
# mid_y -= (x1 - x0) * 0.1
# # Add points for curved line
# edge_x.extend([x0, mid_x, x1, None])
# edge_y.extend([y0, mid_y, y1, None])
# edge_text.append(edge[2]['relationship'])
# edge_trace = go.Scatter(
# x=edge_x,
# y=edge_y,
# line=dict(width=1.5, color='#9CA3AF'),
# hoverinfo='text',
# text=edge_text,
# mode='lines'
# )
# # Create node trace
# node_x = []
# node_y = []
# for node in G.nodes():
# x, y = pos[node]
# node_x.append(x)
# node_y.append(y)
# node_trace = go.Scatter(
# x=node_x,
# y=node_y,
# mode='markers+text',
# hoverinfo='text',
# text=node_texts,
# hovertext=node_hovers, # Full text shown on hover
# textposition="top center",
# marker=dict(
# size=30,
# color=node_colors,
# line=dict(width=2, color='white'),
# symbol='circle'
# ),
# customdata=node_hovers # Store full text for click events
# )
# # Create figure with updated layout
# fig = go.Figure(
# data=[edge_trace, node_trace],
# layout=go.Layout(
# title={
# 'text': 'Document Similarity Graph
Click nodes to view full text',
# 'y': 0.95,
# 'x': 0.5,
# 'xanchor': 'center',
# 'yanchor': 'top'
# },
# showlegend=False,
# hovermode='closest',
# margin=dict(b=20, l=5, r=5, t=60),
# xaxis=dict(showgrid=False, zeroline=False, showticklabels=False),
# yaxis=dict(showgrid=False, zeroline=False, showticklabels=False),
# plot_bgcolor='white',
# width=800,
# height=600,
# clickmode='event+select' # Enable click events
# )
# )
# return fig
# def create_interface(search_system: EnhancedLegalSearchSystem):
# """Create Gradio interface with interactive graph"""
# with gr.Blocks(css="footer {display: none !important;}") as demo:
# gr.Markdown("""
# # Enhanced Legal Search System
# Enter your legal query below to search through documents and get an AI-powered analysis.
# This system combines graph-based and semantic search capabilities for comprehensive legal research.
# """)
# with gr.Row():
# query_input = gr.Textbox(
# label="Legal Query",
# placeholder="e.g., What are the reporting obligations for banks under the Money Laundering Act?",
# lines=3
# )
# with gr.Row():
# search_button = gr.Button("Search & Analyze")
# status_output = gr.Textbox(
# label="Status",
# interactive=False
# )
# with gr.Tabs():
# with gr.TabItem("AI Legal Analysis"):
# analysis_output = gr.Markdown(
# label="AI-Generated Legal Analysis",
# value="Analysis will appear here..."
# )
# with gr.TabItem("Retrieved Documents"):
# docs_output = gr.Markdown(
# label="Source Documents",
# value="Search results will appear here..."
# )
# with gr.TabItem("Related Concepts"):
# concepts_output = gr.Json(
# label="Related Legal Concepts",
# value={}
# )
# with gr.TabItem("Knowledge Graph"):
# # Graph visualization
# graph_output = gr.Plot(
# label="Legal Knowledge Graph"
# )
# # Add text area for showing clicked document content
# selected_doc_content = gr.Textbox(
# label="Selected Document Content",
# interactive=False,
# lines=10
# )
# async def process_query(query):
# if not query.strip():
# return (
# "Please enter a query",
# "No analysis available",
# "No documents available",
# {},
# None,
# ""
# )
# results = await search_system.process_legal_query(query)
# graph_data = search_system.generate_document_graph(query)
# graph_fig = create_graph_visualization(graph_data)
# return (
# results['status'],
# results['answer'],
# results['documents'],
# {"related_concepts": results['related_concepts']},
# graph_fig,
# "Click on a node to view document content"
# )
# search_button.click(
# fn=process_query,
# inputs=[query_input],
# outputs=[
# status_output,
# analysis_output,
# docs_output,
# concepts_output,
# graph_output,
# selected_doc_content
# ]
# )
# return demo
# class LegalSearchSystem:
# def __init__(
# self,
# pinecone_api_key: str = "pcsk_43sajZ_MjcXR2yN5cAcVi8RARyB6i3NP3wLTnTLugbUcN9cUU4q5EfNmuwLPkmxAvykk9o",
# google_api_key: str = "AIzaSyDWHGMd8a70RbL3EBenfUwimcAHjhvgM6M",
# environment: str = "us-east-1",
# index_name: str = "pdf-embeddings",
# dimension: int = 384,
# embedding_model_name: str = "intfloat/e5-small-v2",
# device: str = "cpu"
# ):
# # Initialize Pinecone
# self.pc = Pinecone(api_key=pinecone_api_key)
# # Initialize LangChain with Gemini
# self.llm = ChatGoogleGenerativeAI(
# model="gemini-pro",
# temperature=0,
# google_api_key=google_api_key
# )
# # Initialize prompts
# self.map_prompt = PromptTemplate(
# template="""
# Analyze the following legal text segment and extract key information:
# TEXT: "{text}"
# Instructions:
# 1. Maintain all legal terminology exactly as written
# 2. Preserve section numbers and references
# 3. Keep all specific conditions and requirements
# 4. Include any mentioned time periods or deadlines
# DETAILED ANALYSIS:
# """,
# input_variables=["text"]
# )
# self.combine_prompt = PromptTemplate(
# template="""
# Based on the following excerpts from legal documents and the question: "{question}"
# EXCERPTS:
# {text}
# Instructions:
# 1. Synthesize a comprehensive answer that connects relevant sections
# 2. Maintain precise legal language from the source material
# 3. Reference specific sections and subsections where applicable
# 4. If there are seemingly disconnected pieces of information, explain their relationship
# 5. Highlight any conditions or exceptions that span multiple excerpts
# COMPREHENSIVE LEGAL ANALYSIS:
# """,
# input_variables=["text", "question"]
# )
# # Initialize chain
# self.chain = load_summarize_chain(
# llm=self.llm,
# chain_type="stuff",
# prompt=self.combine_prompt,
# verbose=True
# )
# # Initialize Pinecone index and embedding model
# self.index = self.pc.Index(index_name)
# self.embedding_model = SentenceTransformer(
# model_name_or_path=embedding_model_name,
# device=device
# )
# def search(self, query_text: str, top_k: int = 5, context_window: int = 1) -> Dict:
# """
# Perform a search and analysis of the legal query.
# """
# try:
# # Get search results with context
# results = self.query_and_summarize(
# query_text=query_text,
# top_k=top_k,
# context_window=context_window
# )
# # Format the results for display
# docs_markdown = self._format_documents(results['raw_results'])
# return {
# 'status': "Search completed successfully",
# 'documents': docs_markdown,
# 'analysis': results['summary'],
# 'source_pages': results['source_pages'],
# 'context_info': results['context_info']
# }
# except Exception as e:
# return {
# 'status': f"Error during search: {str(e)}",
# 'documents': "Error retrieving documents",
# 'analysis': "Error generating analysis",
# 'source_pages': [],
# 'context_info': {}
# }
# def query_and_summarize(
# self,
# query_text: str,
# top_k: int = 5,
# filter: Optional[Dict] = None,
# context_window: int = 1
# ) -> Dict:
# """
# Query Pinecone and generate a summary with enhanced context handling.
# """
# # Generate embedding for query
# query_embedding = self.embedding_model.encode(query_text).tolist()
# # Query Pinecone
# initial_results = self.index.query(
# vector=query_embedding,
# top_k=top_k,
# include_metadata=True,
# filter=filter
# )['matches']
# # Expand context
# expanded_results = []
# for match in initial_results:
# page_num = match['metadata']['page_number']
# context_filter = {
# "page_number": {
# "$gte": max(1, page_num - context_window),
# "$lte": page_num + context_window
# }
# }
# if filter:
# context_filter.update(filter)
# context_results = self.index.query(
# vector=self.embedding_model.encode(match['metadata']['text']).tolist(),
# top_k=2 * context_window + 1,
# include_metadata=True,
# filter=context_filter
# )['matches']
# expanded_results.extend(context_results)
# # Process results and generate summary
# documents = self._process_results(expanded_results, initial_results)
# summary = self.chain.run(
# input_documents=documents,
# question=query_text
# )
# return {
# 'raw_results': expanded_results,
# 'summary': summary,
# 'source_pages': list(set(doc.metadata['page_number'] for doc in documents)),
# 'context_info': {
# 'direct_matches': len([d for d in documents if d.metadata['context_type'] == "DIRECT MATCH"]),
# 'context_chunks': len([d for d in documents if d.metadata['context_type'] == "CONTEXT"])
# }
# }
# def _process_results(self, expanded_results: List[Dict], initial_results: List[Dict]) -> List[Document]:
# """
# Process and deduplicate search results.
# """
# seen_ids = set()
# documents = []
# for result in expanded_results:
# if result['id'] not in seen_ids:
# seen_ids.add(result['id'])
# is_direct_match = any(r['id'] == result['id'] for r in initial_results)
# documents.append(Document(
# page_content=result['metadata']['text'],
# metadata={
# 'score': result['score'],
# 'page_number': result['metadata']['page_number'],
# 'context_type': "DIRECT MATCH" if is_direct_match else "CONTEXT"
# }
# ))
# return sorted(documents, key=lambda x: x.metadata['page_number'])
# def _format_documents(self, results: List[Dict]) -> str:
# """
# Format search results as markdown.
# """
# markdown = "### Retrieved Documents\n\n"
# for i, result in enumerate(results, 1):
# markdown += f"**Document {i}** (Page {result['metadata']['page_number']})\n"
# markdown += f"```\n{result['metadata']['text']}\n```\n\n"
# return markdown
# async def process_query_async(query: str, search_system: LegalSearchSystem, graph_search_system: EnhancedLegalSearchSystem):
# """
# Asynchronous function to process both traditional and graph-based searches
# """
# if not query.strip():
# return "Please enter a query", "", "", "", {}
# # Regular search (synchronous)
# results = search_system.search(query)
# try:
# # Graph search (asynchronous)
# graph_results = await graph_search_system.process_legal_query(query)
# graph_documents = graph_results.get('documents', "Error processing graph search")
# graph_concepts = graph_results.get('related_concepts', {})
# except Exception as e:
# graph_documents = f"Error processing graph search: {str(e)}"
# graph_concepts = {}
# graph_data = graph_search_system.generate_document_graph(query)
# graph_fig = create_graph_visualization(graph_data)
# return (
# results['status'],
# results['documents'],
# results['analysis'],
# graph_documents,
# graph_concepts,
# graph_fig,
# "Click on a node to view document content"
# )
# def create_interface(graph_search_system: EnhancedLegalSearchSystem):
# search_system = LegalSearchSystem()
# with gr.Blocks(css="footer {display: none !important;}") as demo:
# gr.Markdown("""
# # Corporate Law Legal Search Engine
# Enter your legal query below to search through documents and get an AI-powered analysis.Queires only related to corporate law will give relevant information
# """)
# with gr.Row():
# query_input = gr.Textbox(
# label="Legal Query",
# placeholder="e.g., What are the key principles of contract law?",
# lines=3
# )
# with gr.Row():
# search_button = gr.Button("Search & Analyze")
# status_output = gr.Textbox(
# label="Status",
# interactive=False
# )
# with gr.Tabs():
# with gr.TabItem("Search Results"):
# docs_output = gr.Markdown(
# label="Retrieved Documents",
# value="Search results will appear here..."
# )
# with gr.TabItem("AI Legal Analysis"):
# summary_output = gr.Markdown(
# label="AI-Generated Legal Analysis",
# value="Analysis will appear here..."
# )
# with gr.TabItem("Retrieved Documents through Graph Rag"):
# docs_output_graph = gr.Markdown(
# label="Source Documents",
# value="Search results will appear here..."
# )
# graph_analysis_output = gr.JSON(
# label="Related Concepts",
# value={}
# )
# with gr.TabItem("Knowledge Graph"):
# # Graph visualization
# graph_output = gr.Plot(
# label="Legal Knowledge Graph"
# )
# # Add text area for showing clicked document content
# selected_doc_content = gr.Textbox(
# label="Selected Document Content",
# interactive=False,
# lines=10
# )
# def process_query(query):
# # Create event loop if it doesn't exist
# try:
# loop = asyncio.get_event_loop()
# except RuntimeError:
# loop = asyncio.new_event_loop()
# asyncio.set_event_loop(loop)
# # Run the async function and get results
# return loop.run_until_complete(
# process_query_async(query, search_system, graph_search_system)
# )
# search_button.click(
# fn=process_query,
# inputs=[query_input],
# outputs=[
# status_output,
# docs_output,
# summary_output,
# docs_output_graph,
# graph_analysis_output,
# graph_output,
# selected_doc_content
# ]
# )
# return demo
# if __name__ == "__main__":
# graph_search_system = EnhancedLegalSearchSystem(
# google_api_key="AIzaSyDWHGMd8a70RbL3EBenfUwimcAHjhvgM6M",
# neo4j_url="neo4j+s://a63462d0.databases.neo4j.io",
# neo4j_username="neo4j",
# neo4j_password="nU8Ut5N8k7LmQzNPe7vUbpZxMirK8rHrmLuzPc2G_Zc"
# )
# demo = create_interface(graph_search_system)
# demo.launch()
import os
import gradio as gr
from pinecone import Pinecone
from sentence_transformers import SentenceTransformer
from typing import List, Dict, Optional
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain.chains.summarize import load_summarize_chain
from langchain.prompts import PromptTemplate, ChatPromptTemplate
from langchain.docstore.document import Document
import time
import asyncio
import plotly.graph_objects as go
from neo4j import GraphDatabase
import networkx as nx
from langchain_community.vectorstores import Neo4jVector
from langchain.chains.summarize import load_summarize_chain
from langchain.chains import LLMChain
from langchain_google_genai import GoogleGenerativeAI, GoogleGenerativeAIEmbeddings
class EnhancedLegalSearchSystem:
def __init__(
self,
google_api_key: str,
neo4j_url: str,
neo4j_username: str,
neo4j_password: str,
embedding_model_name: str = "intfloat/e5-small-v2",
device: str = "cpu"
):
"""Initialize the Enhanced Legal Search System"""
# Initialize LLM
self.llm = GoogleGenerativeAI(
model="gemini-pro",
google_api_key=google_api_key,
temperature=0.1
)
# Initialize embeddings
self.embeddings = GoogleGenerativeAIEmbeddings(
model="models/embedding-001",
google_api_key=google_api_key,
task_type="retrieval_query"
)
# Initialize Neo4j connection
self.neo4j_driver = GraphDatabase.driver(
neo4j_url,
auth=(neo4j_username, neo4j_password)
)
# Initialize vector store
self.vector_store = Neo4jVector.from_existing_graph(
embedding=self.embeddings,
url=neo4j_url,
username=neo4j_username,
password=neo4j_password,
node_label="Document",
text_node_properties=["text"],
embedding_node_property="embedding"
)
# Initialize additional embedding model for enhanced search
self.local_embedding_model = SentenceTransformer(
model_name_or_path=embedding_model_name,
device=device
)
# Initialize prompts
self.init_prompts()
def __del__(self):
"""Cleanup Neo4j connection"""
if hasattr(self, 'neo4j_driver'):
self.neo4j_driver.close()
def init_prompts(self):
"""Initialize enhanced prompts for legal analysis"""
self.qa_prompt = ChatPromptTemplate.from_messages([
("system", """You are a legal expert assistant specializing in Indian law.
Analyze the following legal context and provide a detailed, structured answer to the question.
Include specific sections, rules, and precedents where applicable.
Format your response with clear headings and bullet points for better readability.
Context: {context}"""),
("human", "Question: {question}")
])
self.map_prompt = PromptTemplate(
template="""
Analyze the following legal text segment:
TEXT: "{text}"
Instructions:
1. Extract and summarize the key legal points
2. Maintain all legal terminology exactly as written
3. Preserve section numbers and references
4. Keep all specific conditions and requirements
5. Include any mentioned time periods or deadlines
KEY POINTS:
""",
input_variables=["text"] # Removed page_number as it's not used in the template
)
self.combine_prompt = PromptTemplate(
template="""
Question: {question}
Using ONLY the information from the following legal document excerpts, provide a comprehensive answer:
{text}
Instructions:
1. Base your response EXCLUSIVELY on the provided document excerpts
2. If the documents don't contain enough information to fully answer the question, explicitly state what's missing
3. Use direct quotes when appropriate
4. Organize the response by relevant sections found in the documents
5. If there are conflicting statements across documents, highlight them
ANALYSIS:
""",
input_variables=["text", "question"]
)
# Initialize summarize chain
self.chain = load_summarize_chain(
llm=self.llm,
chain_type="map_reduce",
map_prompt=self.map_prompt,
combine_prompt=self.combine_prompt,
verbose=True
)
def get_related_legal_entities(self, query: str) -> List[Dict]:
"""Retrieve related legal entities and their relationships"""
# Corrected Cypher query to handle aggregation properly
cypher_query = """
// First, let's check if nodes exist and get their labels
MATCH (d:Document)
WHERE toLower(d.text) CONTAINS toLower($query)
WITH d
// Match all relationships from the document, collecting their types
OPTIONAL MATCH (d)-[r]-(connected)
WHERE NOT connected:Document // Avoid direct document-to-document relations
WITH d,
collect(DISTINCT type(r)) as relationTypes,
collect(DISTINCT labels(connected)) as connectedLabels
// Now use these to build our main query
MATCH (d:Document)-[r1]-(e)
WHERE toLower(d.text) CONTAINS toLower($query)
AND NOT e:Document // Exclude direct document connections
WITH d, r1, e
// Get secondary connections, but be more specific about what we're looking for
OPTIONAL MATCH (e)-[r2]-(related)
WHERE (related:Entity OR related:Concept OR related:Section OR related:Case)
AND related <> d // Prevent cycles back to original document
WITH d, {
source_id: id(d),
source_text: d.text,
document_type: COALESCE(d.type, "Unknown"),
relationship_type: type(r1),
entity: {
id: id(e),
type: CASE WHEN e:Entity THEN "Entity"
WHEN e:Concept THEN "Concept"
WHEN e:Section THEN "Section"
WHEN e:Case THEN "Case"
ELSE "Other" END,
text: COALESCE(e.text, e.name, e.title, "Unnamed"),
properties: properties(e)
},
related_entities: collect(DISTINCT {
id: id(related),
type: CASE WHEN related:Entity THEN "Entity"
WHEN related:Concept THEN "Concept"
WHEN related:Section THEN "Section"
WHEN related:Case THEN "Case"
ELSE "Other" END,
relationship: type(r2),
text: COALESCE(related.text, related.name, related.title, "Unnamed"),
properties: properties(related)
})
} as result
WHERE result.entity.text IS NOT NULL // Filter out any results with null entity text
RETURN DISTINCT result
ORDER BY result.source_id, result.entity.id
LIMIT 25
"""
try:
with self.neo4j_driver.session() as session:
# Execute the improved query
result = session.run(cypher_query, {"query": query})
entities = [record["result"] for record in result]
# Log the results for debugging
print(f"Found {len(entities)} related entities")
if entities:
for entity in entities:
print(f"Entity: {entity['entity']['text']}")
print(f"Source: {entity['source_text'][:100]}...")
print(f"Related: {len(entity['related_entities'])} connections")
return entities
except Exception as e:
print(f"Error in get_related_legal_entities: {str(e)}")
return []
async def process_legal_query(
self,
question: str,
top_k: int = 5,
context_window: int = 1
) -> Dict[str, any]:
"""Process a legal query using both graph and vector search capabilities"""
try:
# 1. Perform semantic search
semantic_results = self.vector_store.similarity_search(
question,
k=top_k,
search_type="hybrid"
)
# 2. Get related legal entities with the full question context
related_entities = self.get_related_legal_entities(question)
# Log the counts for debugging
print(f"Found {len(semantic_results)} semantic results")
print(f"Found {len(related_entities)} related entities")
# 3. Expand context with related documents
expanded_results = self.expand_context(
semantic_results,
context_window
)
# 4. Generate comprehensive answer
documents = self._process_results(expanded_results, semantic_results)
# 5. Prepare context for LLM
context = self._prepare_context(documents, related_entities)
# 6. Generate answer using LLM
chain = LLMChain(llm=self.llm, prompt=self.qa_prompt)
response = await chain.ainvoke({
"context": context,
"question": question
})
answer = response.get('text', '')
# 7. Return structured response with explicit related concepts
return {
"status": "Success",
"answer": answer,
"documents": self._format_documents(documents),
"related_concepts": related_entities, # This should now contain data
"source_ids": sorted(list(set(doc.metadata.get('document_id', 'unknown') for doc in documents))),
"context_info": {
"direct_matches": len([d for d in documents if d.metadata.get('context_type') == "DIRECT MATCH"]),
"context_chunks": len([d for d in documents if d.metadata.get('context_type') == "CONTEXT"])
}
}
except Exception as e:
print(f"Error in process_legal_query: {str(e)}") # Add error logging
return {
"status": f"Error: {str(e)}",
"answer": "An error occurred while processing your query.",
"documents": "",
"related_concepts": [],
"source_ids": [],
"context_info": {}
}
def expand_context(
self,
initial_results: List[Document],
context_window: int
) -> List[Document]:
"""Expand context around search results"""
expanded_results = []
seen_ids = set()
for doc in initial_results:
doc_id = doc.metadata.get('document_id', doc.page_content[:50])
if doc_id not in seen_ids:
# Query for related documents
context_results = self.vector_store.similarity_search(
doc.page_content,
k=2 * context_window + 1,
search_type="hybrid"
)
for result in context_results:
result_id = result.metadata.get('document_id', result.page_content[:50])
if result_id not in seen_ids:
expanded_results.append(result)
seen_ids.add(result_id)
return expanded_results
def _process_results(self, expanded_results: List[Document], initial_results: List[Document]) -> List[Document]:
"""Process and deduplicate search results"""
seen_ids = set()
documents = []
for doc in expanded_results:
doc_id = doc.metadata.get('document_id', doc.page_content[:50])
if doc_id not in seen_ids:
seen_ids.add(doc_id)
is_direct_match = any(
r.metadata.get('document_id', r.page_content[:50]) == doc_id
for r in initial_results
)
doc.metadata['context_type'] = (
"DIRECT MATCH" if is_direct_match else "CONTEXT"
)
documents.append(doc)
return sorted(
documents,
key=lambda x: x.metadata.get('document_id', 'unknown')
)
def _prepare_context(
self,
documents: List[Document],
related_entities: List[Dict]
) -> str:
"""Prepare context for LLM processing"""
context = "\n\nLegal Documents:\n" + "\n".join([
f"[Document ID: {doc.metadata.get('document_id', 'unknown')}] {doc.page_content}"
for doc in documents
])
if related_entities:
context += "\n\nRelated Legal Concepts and Relationships:\n"
for entity in related_entities:
context += f"\n• {entity.get('entity', '')}"
if entity.get('related_entities'):
for related in entity['related_entities']:
if related.get('entity'):
context += f"\n - {related['type']}: {related['entity']}"
return context
def _format_documents(self, documents: List[Document]) -> str:
"""Format documents as markdown"""
markdown = "### Retrieved Documents\n\n"
for i, doc in enumerate(documents, 1):
markdown += (
f"**Document {i}** "
f"(ID: {doc.metadata.get('document_id', 'unknown')}, "
f"{doc.metadata.get('context_type', 'UNKNOWN')})\n"
f"```\n{doc.page_content}\n```\n\n"
)
return markdown
def generate_document_graph(
self,
query: str,
top_k: int = 5,
similarity_threshold: float = 0.5
) -> List[Dict]:
"""Generate graph data based on document similarity and relationships"""
try:
# 1. Get initial semantic search results
semantic_results = self.vector_store.similarity_search(
query,
k=top_k,
search_type="hybrid"
)
# 2. Get embeddings for all documents
doc_texts = [doc.page_content for doc in semantic_results]
doc_embeddings = self.local_embedding_model.encode(doc_texts)
# 3. Create graph data structure
graph_data = []
seen_docs = set()
# First, add all documents as nodes
for i, doc in enumerate(semantic_results):
doc_id = doc.metadata.get('document_id', f'doc_{i}')
if doc_id not in seen_docs:
seen_docs.add(doc_id)
doc_type = doc.metadata.get('type', 'document')
# Create node entry
graph_data.append({
'source_id': doc_id,
'source_text': doc.page_content[:200], # Truncate for display
'document_type': doc_type,
'entity': {
'id': doc_id,
'type': 'Document',
'text': f"Document {i + 1}",
'properties': {
'similarity': 1.0,
'length': len(doc.page_content)
}
},
'related_entities': []
})
# Add relationships based on similarity
from sklearn.metrics.pairwise import cosine_similarity
similarity_matrix = cosine_similarity(doc_embeddings)
# Create relationships between similar documents
for i in range(len(semantic_results)):
related = []
for j in range(len(semantic_results)):
if i != j and similarity_matrix[i][j] > similarity_threshold:
doc_j = semantic_results[j]
doc_j_id = doc_j.metadata.get('document_id', f'doc_{j}')
related.append({
'id': doc_j_id,
'type': 'Document',
'relationship': 'similar_to',
'text': f"Document {j + 1}",
'properties': {
'similarity_score': float(similarity_matrix[i][j])
}
})
# Add related documents to the graph data
if related:
graph_data[i]['related_entities'] = related
return graph_data
except Exception as e:
print(f"Error generating document graph: {str(e)}")
return []
def create_graph_visualization(graph_data: List[Dict]) -> go.Figure:
"""Create an interactive graph visualization using Plotly"""
if not graph_data:
return go.Figure(layout=go.Layout(title='No documents found'))
# Initialize graph
G = nx.Graph()
# Color mapping
color_map = {
'Document': '#3B82F6', # blue
'Section': '#10B981', # green
'Reference': '#F59E0B' # yellow
}
# Node information storage
node_colors = []
node_texts = []
node_hovers = [] # Full text for hover
nodes_added = set()
# Process nodes and edges
for data in graph_data:
source_id = data['source_id']
source_text = data['source_text']
# Add main document node
if source_id not in nodes_added:
G.add_node(source_id)
node_colors.append(color_map['Document'])
# Short text for display
node_texts.append(f"Doc {len(nodes_added)+1}")
# Full text for hover/click
node_hovers.append(f"Document {len(nodes_added)+1}:
{source_text}")
nodes_added.add(source_id)
# Process related documents
for related in data.get('related_entities', []):
related_id = related['id']
similarity = related['properties'].get('similarity_score', 0.0)
if related_id not in nodes_added:
G.add_node(related_id)
node_colors.append(color_map['Document'])
node_texts.append(f"Doc {len(nodes_added)+1}")
node_hovers.append(f"Document {len(nodes_added)+1}:
{related['text']}")
nodes_added.add(related_id)
# Add edge with similarity weight
G.add_edge(
source_id,
related_id,
weight=similarity,
relationship=f"Similarity: {similarity:.2f}"
)
# Create layout
pos = nx.spring_layout(G, k=2.0, iterations=50)
# Create edge trace
edge_x = []
edge_y = []
edge_text = []
for edge in G.edges(data=True):
x0, y0 = pos[edge[0]]
x1, y1 = pos[edge[1]]
# Create curved line
mid_x = (x0 + x1) / 2
mid_y = (y0 + y1) / 2
# Add some curvature
mid_x += (y1 - y0) * 0.1
mid_y -= (x1 - x0) * 0.1
# Add points for curved line
edge_x.extend([x0, mid_x, x1, None])
edge_y.extend([y0, mid_y, y1, None])
edge_text.append(edge[2]['relationship'])
edge_trace = go.Scatter(
x=edge_x,
y=edge_y,
line=dict(width=1.5, color='#9CA3AF'),
hoverinfo='text',
text=edge_text,
mode='lines'
)
# Create node trace
node_x = []
node_y = []
for node in G.nodes():
x, y = pos[node]
node_x.append(x)
node_y.append(y)
node_trace = go.Scatter(
x=node_x,
y=node_y,
mode='markers+text',
hoverinfo='text',
text=node_texts,
hovertext=node_hovers, # Full text shown on hover
textposition="top center",
marker=dict(
size=30,
color=node_colors,
line=dict(width=2, color='white'),
symbol='circle'
),
customdata=node_hovers # Store full text for click events
)
# Create figure with updated layout
fig = go.Figure(
data=[edge_trace, node_trace],
layout=go.Layout(
title={
'text': 'Document Similarity Graph
Click nodes to view full text',
'y': 0.95,
'x': 0.5,
'xanchor': 'center',
'yanchor': 'top'
},
showlegend=False,
hovermode='closest',
margin=dict(b=20, l=5, r=5, t=60),
xaxis=dict(showgrid=False, zeroline=False, showticklabels=False),
yaxis=dict(showgrid=False, zeroline=False, showticklabels=False),
plot_bgcolor='white',
width=800,
height=600,
clickmode='event+select' # Enable click events
)
)
return fig
def create_interface(graph_search_system: EnhancedLegalSearchSystem):
"""Create Gradio interface with interactive graph"""
search_system = LegalSearchSystem()
with gr.Blocks(css="footer {display: none !important;}") as demo:
gr.Markdown("""
# Corporate Law Legal Search Engine
Enter your legal query below to search through documents and get an AI-powered analysis. Queries only related to corporate law will give relevant information.
""")
with gr.Row():
query_input = gr.Textbox(
label="Legal Query",
placeholder="e.g., What are the key principles of contract law?",
lines=3
)
with gr.Row():
search_button = gr.Button("Search & Analyze")
status_output = gr.Textbox(
label="Status",
interactive=False
)
with gr.Tabs():
with gr.TabItem("Search Results"):
docs_output = gr.Markdown(
label="Retrieved Documents",
value="Search results will appear here..."
)
with gr.TabItem("AI Legal Analysis"):
summary_output = gr.Markdown(
label="AI-Generated Legal Analysis",
value="Analysis will appear here..."
)
with gr.TabItem("Retrieved Documents through Graph Rag"):
docs_output_graph = gr.Markdown(
label="Source Documents",
value="Search results will appear here..."
)
graph_analysis_output = gr.JSON(
label="Related Concepts",
value={}
)
with gr.TabItem("Knowledge Graph"):
# Graph visualization
graph_output = gr.Plot(
label="Legal Knowledge Graph"
)
# Add text area for showing clicked document content
selected_doc_content = gr.Textbox(
label="Selected Document Content",
interactive=False,
lines=10
)
async def process_query(query):
if not query.strip():
return (
"Please enter a query",
"No documents available from Pinecone",
"No analysis available from Pinecone",
"No documents available from Neo4j",
{},
None,
""
)
# Run the regular RAG search
rag_results = search_system.search(query)
# Run the graph-based RAG search
graph_results = await graph_search_system.process_legal_query(query)
graph_data = graph_search_system.generate_document_graph(query)
graph_fig = create_graph_visualization(graph_data)
return (
rag_results['status'],
rag_results['documents'],
rag_results['analysis'],
graph_results['documents'],
{"related_concepts": graph_results['related_concepts']},
graph_fig,
"Click on a node to view document content"
)
search_button.click(
fn=process_query,
inputs=[query_input],
outputs=[
status_output,
docs_output,
summary_output,
docs_output_graph,
graph_analysis_output,
graph_output,
selected_doc_content
]
)
return demo
class LegalSearchSystem:
def __init__(
self,
pinecone_api_key: str = "pcsk_43sajZ_MjcXR2yN5cAcVi8RARyB6i3NP3wLTnTLugbUcN9cUU4q5EfNmuwLPkmxAvykk9o",
google_api_key: str = "AIzaSyDWHGMd8a70RbL3EBenfUwimcAHjhvgM6M",
environment: str = "us-east-1",
index_name: str = "pdf-embeddings",
dimension: int = 384,
embedding_model_name: str = "intfloat/e5-small-v2",
device: str = "cpu"
):
# Initialize Pinecone
self.pc = Pinecone(api_key=pinecone_api_key)
# Initialize LangChain with Gemini
self.llm = ChatGoogleGenerativeAI(
model="gemini-pro",
temperature=0,
google_api_key=google_api_key
)
# Initialize prompts
self.map_prompt = PromptTemplate(
template="""
Analyze the following legal text segment and extract key information:
TEXT: "{text}"
Instructions:
1. Maintain all legal terminology exactly as written
2. Preserve section numbers and references
3. Keep all specific conditions and requirements
4. Include any mentioned time periods or deadlines
DETAILED ANALYSIS:
""",
input_variables=["text"]
)
self.combine_prompt = PromptTemplate(
template="""
Based on the following excerpts from legal documents and the question: "{question}"
EXCERPTS:
{text}
Instructions:
1. Synthesize a comprehensive answer that connects relevant sections
2. Maintain precise legal language from the source material
3. Reference specific sections and subsections where applicable
4. If there are seemingly disconnected pieces of information, explain their relationship
5. Highlight any conditions or exceptions that span multiple excerpts
COMPREHENSIVE LEGAL ANALYSIS:
""",
input_variables=["text", "question"]
)
# Initialize chain
self.chain = load_summarize_chain(
llm=self.llm,
chain_type="stuff",
prompt=self.combine_prompt,
verbose=True
)
# Initialize Pinecone index and embedding model
self.index = self.pc.Index(index_name)
self.embedding_model = SentenceTransformer(
model_name_or_path=embedding_model_name,
device=device
)
def search(self, query_text: str, top_k: int = 5, context_window: int = 1) -> Dict:
"""
Perform a search and analysis of the legal query.
"""
try:
# Get search results with context
results = self.query_and_summarize(
query_text=query_text,
top_k=top_k,
context_window=context_window
)
# Format the results for display
docs_markdown = self._format_documents(results['raw_results'])
return {
'status': "Search completed successfully",
'documents': docs_markdown,
'analysis': results['summary'],
'source_pages': results['source_pages'],
'context_info': results['context_info']
}
except Exception as e:
return {
'status': f"Error during search: {str(e)}",
'documents': "Error retrieving documents",
'analysis': "Error generating analysis",
'source_pages': [],
'context_info': {}
}
def query_and_summarize(
self,
query_text: str,
top_k: int = 5,
filter: Optional[Dict] = None,
context_window: int = 1
) -> Dict:
"""
Query Pinecone and generate a summary with enhanced context handling.
"""
# Generate embedding for query
query_embedding = self.embedding_model.encode(query_text).tolist()
# Query Pinecone
initial_results = self.index.query(
vector=query_embedding,
top_k=top_k,
include_metadata=True,
filter=filter
)['matches']
# Expand context
expanded_results = []
for match in initial_results:
page_num = match['metadata']['page_number']
context_filter = {
"page_number": {
"$gte": max(1, page_num - context_window),
"$lte": page_num + context_window
}
}
if filter:
context_filter.update(filter)
context_results = self.index.query(
vector=self.embedding_model.encode(match['metadata']['text']).tolist(),
top_k=2 * context_window + 1,
include_metadata=True,
filter=context_filter
)['matches']
expanded_results.extend(context_results)
# Process results and generate summary
documents = self._process_results(expanded_results, initial_results)
summary = self.chain.run(
input_documents=documents,
question=query_text
)
return {
'raw_results': expanded_results,
'summary': summary,
'source_pages': list(set(doc.metadata['page_number'] for doc in documents)),
'context_info': {
'direct_matches': len([d for d in documents if d.metadata['context_type'] == "DIRECT MATCH"]),
'context_chunks': len([d for d in documents if d.metadata['context_type'] == "CONTEXT"])
}
}
def _process_results(self, expanded_results: List[Dict], initial_results: List[Dict]) -> List[Document]:
"""
Process and deduplicate search results.
"""
seen_ids = set()
documents = []
for result in expanded_results:
if result['id'] not in seen_ids:
seen_ids.add(result['id'])
is_direct_match = any(r['id'] == result['id'] for r in initial_results)
documents.append(Document(
page_content=result['metadata']['text'],
metadata={
'score': result['score'],
'page_number': result['metadata']['page_number'],
'context_type': "DIRECT MATCH" if is_direct_match else "CONTEXT"
}
))
return sorted(documents, key=lambda x: x.metadata['page_number'])
def _format_documents(self, results: List[Dict]) -> str:
"""
Format search results as markdown.
"""
markdown = "### Retrieved Documents\n\n"
for i, result in enumerate(results, 1):
markdown += f"**Document {i}** (Page {result['metadata']['page_number']})\n"
markdown += f"```\n{result['metadata']['text']}\n```\n\n"
return markdown
if __name__ == "__main__":
graph_search_system = EnhancedLegalSearchSystem(
google_api_key="AIzaSyDWHGMd8a70RbL3EBenfUwimcAHjhvgM6M",
neo4j_url="neo4j+s://a63462d0.databases.neo4j.io",
neo4j_username="neo4j",
neo4j_password="nU8Ut5N8k7LmQzNPe7vUbpZxMirK8rHrmLuzPc2G_Zc"
)
demo = create_interface(graph_search_system)
demo.launch()