# AlzDetect AI — Streamlit App ───────────
from anyio import Path
import streamlit as st
import faiss
import numpy as np
import json
import os
import anthropic
from sentence_transformers import SentenceTransformer
from dotenv import load_dotenv
# Fix — Load .env from root folder
env_path = Path(__file__).parent.parent / ".env"
load_dotenv(dotenv_path=env_path)
# Verify key loaded
api_key = os.getenv("ANTHROPIC_API_KEY")
if not api_key:
raise ValueError("❌ ANTHROPIC_API_KEY not found!")
# ── Page Config ────────────────────────────
st.set_page_config(
page_title="AlzDetect AI",
page_icon="🧠",
layout="centered",
initial_sidebar_state="expanded"
)
# ── Custom CSS ─────────────────────────────
st.markdown("""
""", unsafe_allow_html=True)
# ── Load Components (Cached) ───────────────
@st.cache_resource
def load_components():
"""Load all RAG components once"""
# Paths
CHUNKS_FILE = 'data/processed/chunks.json'
FAISS_FILE = 'vector_store/faiss_index/alzheimer.index'
# Load chunks
with open(CHUNKS_FILE) as f:
chunks = json.load(f)
# Load FAISS
index = faiss.read_index(FAISS_FILE)
# Load model
model = SentenceTransformer(
'neuml/pubmedbert-base-embeddings'
)
# Load Claude
client = anthropic.Anthropic(
api_key=api_key # ← use variable!
)
return chunks, index, model, client
# ── RAG Functions ──────────────────────────
def retrieve_chunks(query, chunks, index, model, k=5):
query_vector = model.encode([query])
query_vector = query_vector.astype('float32')
distances, indices = index.search(query_vector, k=k)
results = []
for dist, idx in zip(distances[0], indices[0]):
chunk = chunks[idx]
results.append({
"title": chunk['title'],
"authors": chunk['authors'],
"year": chunk['year'],
"journal": chunk['journal'],
"chunk": chunk['chunk'],
"pmid": chunk['pmid'],
"distance": float(dist)
})
return results
def generate_answer(query, retrieved, client):
context = ""
for i, r in enumerate(retrieved):
context += f"""
Source {i+1}:
Title: {r['title']}
Authors: {r['authors']}
Year: {r['year']}
Journal: {r['journal']}
Content: {r['chunk']}
---"""
prompt = f"""You are AlzDetect AI — expert
medical research assistant specializing in
Alzheimer's disease research.
Answer using ONLY these PubMed sources.
Cite as (Source 1), (Source 2) etc.
Be specific with biomarker names and findings.
Use markdown formatting.
Never hallucinate.
SOURCES:
{context}
QUESTION: {query}"""
response = client.messages.create(
model="claude-opus-4-5",
max_tokens=1500,
messages=[{
"role": "user",
"content": prompt
}]
)
return response.content[0].text
# ── Main App ───────────────────────────────
def main():
# Header
st.markdown(
'🧠 AlzDetect AI
',
unsafe_allow_html=True
)
st.markdown(
'',
unsafe_allow_html=True
)
# Load components
with st.spinner("Loading AI components..."):
chunks, index, model, client = load_components()
# Sidebar
with st.sidebar:
st.markdown("## Pipeline Stats")
st.markdown(
'📄 16,281 Papers
',
unsafe_allow_html=True
)
st.markdown("")
st.markdown(
'🧩 40,995 Chunks
',
unsafe_allow_html=True
)
st.markdown("")
st.markdown(
'🧠 768-dim PubMedBERT
',
unsafe_allow_html=True
)
st.markdown("")
st.markdown(
'🔍 FAISS Vector Search
',
unsafe_allow_html=True
)
st.markdown("")
st.markdown(
'🤖 Claude AI
',
unsafe_allow_html=True
)
st.markdown("---")
st.markdown("## 💡 Sample Questions")
sample_questions = [
"What blood biomarkers detect Alzheimer's early?",
"How does diabetes connect to Alzheimer's?",
"How effective is Lecanemab treatment?",
"What is the role of APOE4 in dementia?",
"How does deep learning help diagnose Alzheimer's?"
]
for q in sample_questions:
if st.button(q, key=q):
st.session_state.query = q
st.markdown("---")
st.markdown("## ℹ️ About")
st.markdown("""
AlzDetect AI uses enterprise RAG
to answer Alzheimer's research
questions with full citations.
**Data source:** PubMed
**Model:** PubMedBERT
**LLM:** Claude API
**Built by:** TPriya
""")
# Main query input
query = st.text_input(
"🔍 Ask a research question:",
value=st.session_state.get('query', ''),
placeholder="e.g. What blood biomarkers detect Alzheimer's early?"
)
col1, col2 = st.columns([1, 4])
with col1:
search_btn = st.button(
"🔍 Search",
type="primary",
use_container_width=True
)
if search_btn and query:
# Search
with st.spinner("🔍 Searching 40,995 research chunks..."):
retrieved = retrieve_chunks(
query, chunks, index, model, k=5
)
# Generate
with st.spinner("🤖 Generating cited answer with Claude..."):
answer = generate_answer(
query, retrieved, client
)
# Display answer
st.markdown("### 🤖 Answer")
st.markdown(
f'{answer}
',
unsafe_allow_html=True
)
# Display sources
st.markdown("### 📚 Sources Retrieved")
for i, r in enumerate(retrieved):
with st.expander(
f"Source {i+1}: {r['title'][:60]}..."
):
st.markdown(f"**Title:** {r['title']}")
st.markdown(f"**Authors:** {r['authors']}")
st.markdown(f"**Year:** {r['year']}")
st.markdown(f"**Journal:** {r['journal']}")
st.markdown(f"**PMID:** {r['pmid']}")
st.markdown(f"**Relevance Score:** {r['distance']:.2f}")
st.markdown("**Abstract Excerpt:**")
st.markdown(f"_{r['chunk'][:300]}..._")
# Footer metrics
st.markdown("---")
col1, col2, col3 = st.columns(3)
with col1:
st.metric(
"Sources Found",
len(retrieved)
)
with col2:
st.metric(
"Best Distance",
f"{retrieved[0]['distance']:.2f}"
)
with col3:
st.metric(
"Papers Searched",
"16,281"
)
elif search_btn and not query:
st.warning(
" Please enter a question first!"
)
# Footer
st.markdown("---")
st.markdown(
"""
AlzDetect AI | Built with PubMed × PubMedBERT × FAISS × Claude
| For research purposes only — not medical advice
""",
unsafe_allow_html=True
)
if __name__ == "__main__":
main()