app.py

import os
import streamlit as st
import torch
from datasets import load_dataset
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

# ===================================
# PAGE CONFIG
# ===================================
st.set_page_config(
    page_title="Kenya Legal Assistant",
    layout="wide"
)

st.title("🇰🇪 Kenya Legal Assistant")
st.caption("Ask questions about Kenyan court judgments (English or Swahili)")

# ===================================
# LOAD VECTOR DATABASE (CACHED)
# ===================================
@st.cache_resource(show_spinner=True)
def load_vectorstore():

    st.write("🔎 Loading legal knowledge base...")

    dataset = load_dataset(
        "Brian269/Kenyan_Judgements",
        split="train",
        streaming=True
    )

    documents = []
    for i, item in enumerate(dataset):
        if i > 200:   # prevents HF startup timeout
            break

        documents.append(
            Document(
                page_content=item["text"],
                metadata={
                    "source": item["file_name"],
                    "page": 1
                },
            )
        )

    splitter = RecursiveCharacterTextSplitter(
        chunk_size=1200,
        chunk_overlap=200
    )

    chunks = []
    for doc in documents:
        for chunk in splitter.split_text(doc.page_content):
            chunks.append(
                Document(page_content=chunk, metadata=doc.metadata)
            )

    embeddings = HuggingFaceEmbeddings(
        model_name="sentence-transformers/paraphrase-multilingual-mpnet-base-v2"
    )

    INDEX_PATH = "faiss_index"

    # ✅ Load prebuilt FAISS index if uploaded
    if os.path.exists(INDEX_PATH):
        st.write("✅ Loading FAISS index...")
        vectorstore = FAISS.load_local(
            INDEX_PATH,
            embeddings,
            allow_dangerous_deserialization=True
        )
    else:
        st.warning("⚠️ FAISS index not found — building (first run only)...")
        vectorstore = FAISS.from_documents(chunks, embeddings)
        vectorstore.save_local(INDEX_PATH)

    return vectorstore


# ===================================
# LOAD LANGUAGE MODEL (CACHED)
# ===================================
@st.cache_resource(show_spinner=True)
def load_llm():

    st.write("🧠 Loading language model...")

    model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"

    tokenizer = AutoTokenizer.from_pretrained(model_id)

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        low_cpu_mem_usage=True
    )

    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        max_new_tokens=512,
        temperature=0.2
    )

    return pipe


# Load once
vectorstore = load_vectorstore()
pipe = load_llm()

# ===================================
# HELPERS
# ===================================
def detect_language(text):
    try:
        return detect(text)
    except:
        return "en"


def translate(text, target_lang):
    return GoogleTranslator(source="auto", target=target_lang).translate(text)


def build_prompt(question, context):
    return f"""
You are a Kenyan legal assistant.

Answer ONLY using the provided context.
Include proper case citations.
Do not fabricate information.

Context:
{context}

Question:
{question}

Structured Answer:
"""


def ask_kenya_law(question):

    language = detect_language(question)

    question_en = (
        translate(question, "en") if language == "sw" else question
    )

    retrieved_docs = vectorstore.similarity_search(question_en, k=4)

    context = "\n\n".join([doc.page_content for doc in retrieved_docs])

    prompt = build_prompt(question_en, context)

    result = pipe(prompt)[0]["generated_text"]

    if language == "sw":
        result = translate(result, "sw")

    sources = "\n".join(
        [f'{doc.metadata["source"]} - Page {doc.metadata["page"]}'
         for doc in retrieved_docs]
    )

    return result, sources


# ===================================
# STREAMLIT CHAT UI
# ===================================
if "messages" not in st.session_state:
    st.session_state.messages = []

# Display history
for msg in st.session_state.messages:
    with st.chat_message(msg["role"]):
        st.markdown(msg["content"])

prompt = st.chat_input("Ask a legal question...")

if prompt:

    st.session_state.messages.append(
        {"role": "user", "content": prompt}
    )

    with st.chat_message("user"):
        st.markdown(prompt)

    with st.chat_message("assistant"):
        with st.spinner("Analyzing Kenyan case law..."):
            answer, sources = ask_kenya_law(prompt)

            response = f"""
{answer}

---

📚 **Sources**
{sources}

⚠️ DISCLAIMER:
This AI provides legal information for educational purposes only.
It does NOT constitute legal advice.
"""

            st.markdown(response)

    st.session_state.messages.append(
        {"role": "assistant", "content": response}
    )