Create app.py

app.py

@@ -0,0 +1,136 @@
+import os
+import streamlit as st
+from qdrant_client import QdrantClient
+from langchain_qdrant import QdrantVectorStore, RetrievalMode
+from langchain_huggingface import HuggingFaceEmbeddings
+from sentence_transformers import CrossEncoder
+from langchain_groq import ChatGroq
+
+# ------------------------------
+# Streamlit Config
+# ------------------------------
+st.set_page_config(
+    page_title="Nepal Constitution AI",
+    page_icon="🧑‍⚖️",
+    layout="wide"
+)
+
+st.title("🧑‍⚖️ Nepal Constitution – AI Legal Assistant")
+st.caption("Hybrid RAG + Cross-Encoder Reranking (Demo)")
+
+# ------------------------------
+# User Input
+# ------------------------------
+query = st.text_input(
+    "Ask a constitutional or legal question:",
+    placeholder="e.g. What does Article 275 say about local governance?"
+)
+
+# ------------------------------
+# Cached Models (VERY IMPORTANT)
+# ------------------------------
+@st.cache_resource
+def load_embeddings():
+    return HuggingFaceEmbeddings(
+        model_name="BAAI/bge-m3",
+        model_kwargs={"device": "cpu"},
+        encode_kwargs={"normalize_embeddings": True}
+    )
+
+@st.cache_resource
+def load_reranker():
+    return CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")
+
+@st.cache_resource
+def load_vector_store():
+    client = QdrantClient(path="./qdrant_db")
+    embeddings = load_embeddings()
+
+    return QdrantVectorStore(
+        path = "./qdrant_db",
+        collection_name="nepal_law",
+        embedding=embeddings,
+        retrieval_mode=RetrievalMode.HYBRID
+    )
+
+@st.cache_resource
+def load_llm():
+    return ChatGroq(
+        model="llama-3.1-8b-instant",
+        temperature=0.2,
+        max_tokens=600
+    )
+
+# ------------------------------
+# Reranking Function
+# ------------------------------
+def rerank(query, docs, top_k=6):
+    reranker = load_reranker()
+    pairs = [(query, d.page_content) for d in docs]
+    scores = reranker.predict(pairs)
+
+    ranked = sorted(
+        zip(docs, scores),
+        key=lambda x: x[1],
+        reverse=True
+    )
+
+    return [doc for doc, _ in ranked[:top_k]]
+
+# ------------------------------
+# Main Logic
+# ------------------------------
+if query:
+    with st.spinner("🔍 Searching constitutional knowledge..."):
+        vector_store = load_vector_store()
+
+        # Step 1: Retrieve
+        retrieved_docs = vector_store.similarity_search(query, k=20)
+
+        # Step 2: Rerank
+        reranked_docs = rerank(query, retrieved_docs, top_k=8)
+
+        # Build context
+        context = "\n\n".join(
+            [f"[Source {i+1}]\n{doc.page_content}"
+             for i, doc in enumerate(reranked_docs)]
+        )
+
+    # ------------------------------
+    # Improved Legal Prompt
+    # ------------------------------
+    prompt = f"""
+You are a constitutional law assistant for Nepal.
+
+INSTRUCTIONS:
+- Answer ONLY using the provided context.
+- If the answer is not clearly found in the context, say:
+  "The provided constitutional text does not explicitly answer this question."
+- Do NOT invent articles, clauses, or interpretations.
+- Use clear, formal, and neutral legal language.
+- When relevant, reference article numbers/section numbers mentioned in the context.
+
+CONTEXT:
+{context}
+
+QUESTION:
+{query}
+
+ANSWER:
+"""
+
+    with st.spinner("🧠 Generating answer..."):
+        llm = load_llm()
+        response = llm.invoke(prompt)
+
+    # ------------------------------
+    # Output
+    # ------------------------------
+    st.markdown("### ✅ Answer")
+    st.write(response.content)
+
+    with st.expander("📚 Retrieved Constitutional Sources"):
+        for i, doc in enumerate(reranked_docs):
+            st.markdown(f"**Source {i+1}**")
+            st.write(doc.page_content)
+            st.markdown("---")