Upload 6 files

.env

@@ -0,0 +1,9 @@
+GROQ_API_KEY_1="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"
+GROQ_API_KEY_2="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"
+
+
+#GROQ_API_KEY = "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"
+LANGCHAIN_API_KEY = "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"
+LANGCHAIN_PROJECT = "WORK"
+LANGCHAIN_TRACING_V2 = "true"
+#GOOGLE_API_KEY = "684ff9d7746ed6e9d54268345c5ad41655ffac7d"

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+index.faiss filter=lfs diff=lfs merge=lfs -text

constitution_py.py

@@ -0,0 +1,171 @@
+import warnings
+warnings.filterwarnings("ignore")
+
+import re
+import os
+import numpy as np
+import faiss
+from sentence_transformers import SentenceTransformer
+from langchain_groq import ChatGroq
+from langchain.chains import LLMChain
+from langchain_core.prompts import ChatPromptTemplate
+from pydantic import BaseModel, Field
+from langchain.output_parsers import PydanticOutputParser
+from lm import get_query_llm, get_answer_llm  # Your custom LLM wrapper functions
+from functools import lru_cache
+
+# Initialize LLMs
+q_llm = get_query_llm()
+a_llm = get_answer_llm()
+
+# Load sentence transformer model once globally
+embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
+save_dir = "saved_data"
+
+from functools import lru_cache
+
+# Cache embeddings and index loading
+@lru_cache(maxsize=1)
+def load_embeddings_and_index(save_dir="saved_data"):
+    embedding = np.load(os.path.join(save_dir, "embeddings.npy"))
+    index = faiss.read_index(os.path.join(save_dir, "index.faiss"))
+    with open(os.path.join(save_dir, "chunks.txt"), "r", encoding="utf-8") as f:
+        chunks = [line.strip() for line in f.readlines()]
+    return embedding, index, chunks
+
+
+similar_words = [
+    "explain", "elaborate", "describe", "clarify", "detail", "break down", "simplify", "outline",
+    "demonstrate", "illustrate", "interpret", "expand on", "go over", "walk through", "define",
+    "unpack", "decode", "shed light on", "analyze", "discuss", "make clear", "reveal", "disclose",
+    "comment on", "talk about", "lay out", "spell out", "express", "delve into", "explore",
+    "enlighten", "present", "review", "report", "state", "point out", "inform", "highlight"
+]
+
+def is_explanation_query(query):
+    return not any(word in query.lower() for word in similar_words)
+
+def retrieve_relevant_chunks(query, index, chunks, top_k=5):
+    sub_str = "article"
+    numbers = re.findall(r'\d+', query)
+    flag = False
+    if sub_str in query.lower() and numbers:
+        article_number = str(numbers[0])
+        for i, chunk in enumerate(chunks):
+            if chunk.lower().startswith(f"article;{article_number}"):
+                flag = is_explanation_query(query)
+                
+                return [chunk], flag
+    print(flag)
+    query_embedding = embedding_model.encode([query])
+    query_embedding = np.array(query_embedding).astype("float32")
+    distances, indices = index.search(query_embedding, top_k)
+    relevant_chunks = [chunks[i] for i in indices[0]]
+    return relevant_chunks, flag
+
+# Prompt to refine the query
+refine_prompt_template = ChatPromptTemplate.from_messages([
+    ('system', 
+        "You are a legal assistant specialized in cleaning user queries. "
+        "Your task is to fix spelling mistakes and convert number words to digits only (e.g., 'three' to '3'). "
+        "Do not correct grammar, punctuation, or capitalization. "
+        "Do not restructure or rephrase the query in any way. "
+        "Do not add or remove words. "
+        "If the input is already clean or does not make sense, return it exactly as it is. "
+        "Only return one corrected query."),
+    ('human', '{query}')
+])
+refine_chain = LLMChain(llm=q_llm, prompt=refine_prompt_template)
+
+# Define response schema
+class LegalResponse(BaseModel):
+    title: str = Field (...,description='Return the title')
+    answer: str = Field(..., description="The assistant's answer to the user's query")
+    is_relevant: bool = Field(..., description="True if the query is relevant to the Constitution of Pakistan, otherwise False")
+    article_number: str = Field(..., description="Mentioned article number if available, else empty string")
+
+parser = PydanticOutputParser(pydantic_object=LegalResponse)
+
+# Prompt for direct article wording
+answer_prompt_template_query = ChatPromptTemplate.from_messages([
+    ("system",
+     "You are a legal assistant with expertise in the Constitution of Pakistan. "
+     "Return answer in structure format."
+     "Your task is to extract and present the exact constitutional text, without paraphrasing, ensuring accuracy and fidelity to the original wording"
+     "Especially return the title"),
+    ("human",
+     "User Query: {query}\n\n"
+     "Instructions:\n"
+     "0. Return Title"
+     "1. Return the exact wording from the Constitution.\n"
+     "2. If a query references a specific article or sub-clause (e.g., Article 11(3)(b), Article 11(b), or 11(i)), return only the exact wording of that clause from the Constitution — do not include the full article unless required by structure\n"
+     "3. Indicate whether the query is related to the Constitution of Pakistan (Yes/No).ar\n"
+     "4. Extract and return the article number if it is mentioned. with sub-clause if its mentioned like 1,2 or 1(a)\n\n"
+     "Context:\n{context}\n\n"
+     "{format_instructions}\n")
+])
+
+answer_chain_article = LLMChain(llm=a_llm, prompt=answer_prompt_template_query, output_parser=parser)
+
+# Prompt for explanation-style answers
+explanation_prompt_template_query = ChatPromptTemplate.from_messages([
+    ("system", 
+     "You are a legal expert assistant with deep knowledge of the Constitution of Pakistan. "
+     "You will receive a user query and a set of context chunks from the Constitution. "
+     "Your task is to determine if the query is answerable based strictly on the information provided in the context. "
+     "If it is, provide a structured explanation based on that context—without copying or repeating the context text verbatim. "
+     "If the information needed to answer is not found in the provided chunks, respond with a structured message indicating `Is Relevant: False`, and do not fabricate any information."
+    ),
+    
+    ("human",
+     "User Query: {query}\n\n"
+     "Context (Extracted Chunks):\n{context}\n\n"
+     "Instructions:\n"
+     "1. Use only the information in the context to determine if the query can be answered.\n"
+     "2. DO NOT include or repeat the context text directly in your answer. Summarize or paraphrase when needed.\n"
+     "3. If the query is answerable based on the context, explain the related article, clause, or provision clearly and precisely:\n"
+     "   - Include the Article number if available.\n"
+     "   - Describe its meaning and how it functions within the Constitution.\n"
+     "4. Do NOT use real-world references, court cases, or examples.\n"
+     "5. Conclude your response with:\n"
+     "   - `Is Relevant: True/False`\n"
+     "   - `Related Article(s)`: List article number(s) if any.\n\n"
+     "{format_instructions}\n")
+])
+
+answer_chain_explanation = LLMChain(llm=a_llm, prompt=explanation_prompt_template_query, output_parser=parser)
+
+# Load data
+embeddings, index, chunks = load_embeddings_and_index(save_dir)
+
+# Main function
+def get_legal_response(query):
+    try:
+        refined_query = refine_chain.run(query=query)
+    except Exception as e:
+        print(f"[Refinement Error] Using raw query instead: {e}")
+        refined_query = query
+
+    print("\nRefined Query:", refined_query)
+
+    relevant_chunks, flag = retrieve_relevant_chunks(refined_query, index, chunks, top_k=5)
+
+    print("\nTop Relevant Chunks:")
+    for i, chunk in enumerate(relevant_chunks, 1):
+        print(f"\nChunk {i}:\n{'-'*50}\n{chunk}")
+
+    context = "\n\n".join(relevant_chunks)
+
+    if flag==True:
+        print('okokokokokokokokokokok')
+        response = answer_chain_article.run(query=refined_query,context=context,format_instructions=parser.get_format_instructions())
+    else:
+        print('nononononononononono')
+        response = answer_chain_explanation.run(query=refined_query,context=context,format_instructions=parser.get_format_instructions())
+
+    return {
+        "title":response.title,
+        "answer": response.answer,
+        "is_relevant": response.is_relevant,
+        "article_number": response.article_number
+    }

embeddings.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7594f7c33917c21f7168af5f896d08589c2f9644e6f31096665d0738cf8a7aed
+size 427136

index.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78883b11d6015e1f366dfe10d42b8582b09906c836eba65dd1dd5bf3f8391471
+size 427053

lm.py

@@ -0,0 +1,15 @@
+import os
+from dotenv import load_dotenv
+from langchain_groq import ChatGroq
+
+load_dotenv()
+key1 = os.getenv('GROQ_API_KEY_1')
+key2 = os.getenv('GROQ_API_KEY_2')
+
+
+
+def get_query_llm():
+    return ChatGroq(model="mistral-saba-24b",api_key=key1)
+
+def get_answer_llm():
+    return ChatGroq(model="mistral-saba-24b",api_key=key2)