app.py

import os
from datasets import load_dataset
import torch
from langdetect import detect
from deep_translator import GoogleTranslator
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import FAISS
from langchain.docstore.document import Document
from langchain.embeddings import HuggingFaceEmbeddings
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import gradio as gr

# -----------------------------
# Load PDFs from Hugging Face dataset
# -----------------------------
dataset = load_dataset("Brian269/Kenyan_Judgements", split="train")  # Replace with your dataset

documents = []
for item in dataset:
    pdf_text = item["text"]  # Assuming your dataset has a "text" field
    doc = Document(page_content=pdf_text, metadata={"source": item["file_name"], "page": 1})
    documents.append(doc)

# -----------------------------
# Split text into chunks
# -----------------------------
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1200, chunk_overlap=200)
chunks = []
for doc in documents:
    for chunk in text_splitter.split_text(doc.page_content):
        chunks.append(Document(page_content=chunk, metadata=doc.metadata))

# -----------------------------
# Embeddings + FAISS index
# -----------------------------
embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/paraphrase-multilingual-mpnet-base-v2")
vectorstore = FAISS.from_documents(chunks, embedding_model)

# -----------------------------
# Load LLM
# -----------------------------
model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=512, temperature=0.2)

# -----------------------------
# Helpers for multilingual queries
# -----------------------------
def detect_language(query):
    try:
        return detect(query)
    except:
        return "en"

def translate_text(text, target_lang):
    if target_lang == "sw":
        return GoogleTranslator(source='auto', target='sw').translate(text)
    elif target_lang == "en":
        return GoogleTranslator(source='auto', target='en').translate(text)
    return text

# -----------------------------
# Build prompts
# -----------------------------
DISCLAIMER_TEXT = """
⚠️ DISCLAIMER:
This AI assistant provides legal information derived from publicly
available Kenyan court judgments for educational purposes only.
It does NOT provide legal advice.
For professional legal assistance, consult a qualified advocate.
"""

def build_prompt(question, context):
    instruction = """
You are a Kenyan legal assistant.
Answer concisely using ONLY the provided context.
Include proper case citation (case name and page).
Do not fabricate information.
"""
    return f"{instruction}\n\nContext:\n{context}\n\nQuestion:\n{question}\n\nProvide a clear structured answer."

# -----------------------------
# Query system
# -----------------------------
def ask_kenya_law(question, k=4):
    language = detect_language(question)
    translated_question = translate_text(question, "en") if language == "sw" else question
    retrieved_docs = vectorstore.similarity_search(translated_question, k=k)
    context = "\n\n".join([doc.page_content for doc in retrieved_docs])
    prompt = build_prompt(translated_question, context)
    response = pipe(prompt)[0]["generated_text"]
    if language == "sw":
        response = translate_text(response, "sw")
    sources = [f'{doc.metadata["source"]} - Page {doc.metadata["page"]}' for doc in retrieved_docs]
    return response, "\n".join(sources)

# -----------------------------
# Gradio Interface
# -----------------------------
def query_system(user_input):
    answer, sources = ask_kenya_law(user_input)
    return answer + "\n\n📚 SOURCES:\n" + sources + DISCLAIMER_TEXT

iface = gr.Interface(
    fn=query_system,
    inputs="text",
    outputs="text",
    title="Kenya Legal Assistant",
    description="Ask questions about Kenyan court judgments in English or Swahili."
)

iface.launch()