Update app.py

app.py

@@ -1,4 +1,3 @@
-# app.py
 import os
 import numpy as np
 import streamlit as st
@@ -6,62 +5,31 @@ import faiss
 from sentence_transformers import SentenceTransformer
 from groq import Groq
 
-# ------------------------- Secrets / API Key -------------------------
+# ---------- Secrets / API Key ----------
 API_KEY = os.getenv("GROQ_API_KEY")
 if not API_KEY:
     st.error(
         "GROQ_API_KEY not found.\n\n"
         "Go to your Space → Settings → Repository secrets → Add new secret\n"
         "Name: GROQ_API_KEY | Value: <your Groq key>\n\n"
-        "Then Restart this Space."
+        "Then Restart/Restart this Space."
     )
     st.stop()
 
-# ------------------------- Groq Client -------------------------
+# ---------- Groq Client ----------
 client = Groq(api_key=API_KEY)
 
-# ------------------------- System Prompt (concise) -------------------------
-SYSTEM_MSG = """
-You are a relationship counselor. Read the provided WhatsApp chat context and answer the user’s question.
-Be CONCISE and ACTIONABLE. Do not include internal reasoning or long explanations.
-
-Return Markdown in exactly this format:
-
-Toxicity score: X/10
-
-**1) One-line summary**
-- <max 25 words>
-
-**2) Top red flags (max 3)**
-- <short phrase + why it matters>
-- <short phrase + why it matters>
-- <short phrase + why it matters>
-
-**3) Greens (if any, max 2)**
-- <short phrase>
-- <short phrase>
-
-**4) Next steps (max 3)**
-- <clear, practical action>
-- <clear, practical action>
-- <clear, practical action>
-
-Rules:
-- Keep total answer under 120 words.
-- Use plain language; no therapy jargon.
-- If signs of abuse, add: “**Safety note:** consider talking to a trusted person/professional.”
-"""
-
-# ------------------------- Embeddings / FAISS -------------------------
+# ---------- Models / Index ----------
 @st.cache_resource
 def load_embedder():
-    return SentenceTransformer("all-MiniLM-L6-v2")  # 384-dim
+    # 384-dim embeddings
+    return SentenceTransformer("all-MiniLM-L6-v2")
 
 embedding_model = load_embedder()
-EMB_DIM = 384
 
+DIM = 384  # all-MiniLM-L6-v2 dimension
 if "faiss_index" not in st.session_state:
-    st.session_state.faiss_index = faiss.IndexFlatL2(EMB_DIM)
+    st.session_state.faiss_index = faiss.IndexFlatL2(DIM)
 if "chunks_store" not in st.session_state:
     st.session_state.chunks_store = []
 
@@ -69,8 +37,9 @@ index = st.session_state.faiss_index
 chunks_store = st.session_state.chunks_store
 
 
-# ------------------------- Helpers -------------------------
+# ---------- Helpers ----------
 def chunk_text(text: str, max_length: int = 500):
+    """Simple whitespace chunker by character budget."""
     words, chunks, cur = text.split(), [], []
     for w in words:
         if len(" ".join(cur)) + len(w) + 1 <= max_length:
@@ -86,33 +55,48 @@ def chunk_text(text: str, max_length: int = 500):
 def embed_and_store(chunks):
     if not chunks:
         return
-    embs = embedding_model.encode(chunks, convert_to_numpy=True).astype("float32")
+    embs = embedding_model.encode(
+        chunks, convert_to_numpy=True, normalize_embeddings=False
+    ).astype("float32")
     index.add(embs)
     chunks_store.extend(chunks)
 
 
 def query_llm(prompt: str) -> str:
-    """Non-streaming version for stability on Hugging Face."""
-    try:
-        completion = client.chat.completions.create(
-            model="deepseek-r1-distill-llama-70b",
-            messages=[
-                {"role": "system", "content": SYSTEM_MSG},
-                {"role": "user", "content": prompt},
-            ],
-            temperature=0.5,
-            max_completion_tokens=300,
-            top_p=0.9,
-        )
-        # Return final text directly
-        return completion.choices[0].message.content.strip()
-    except Exception as e:
-        return f"⚠️ **Error:** {str(e)}"
-
-
-# ------------------------- UI -------------------------
+    """Stream a response from Groq and return full text."""
+    stream = client.chat.completions.create(
+        model="deepseek-r1-distill-llama-70b",
+        messages=[
+            {
+                "role": "system",
+                "content": (
+                    "Based on this WhatsApp chat, analyze whether this relationship is healthy or toxic."
+                    "Give a toxicity Score out of 10"
+                    "Highlight top 3 red flags"
+                    "1 positive aspect" 
+                    "3 improvement suggestions"
+                    "Start every answer with: 'Toxicity score: X/10'".
+                    
+                ),
+            },
+            {"role": "user", "content": prompt},
+        ],
+        temperature=0.6,
+        max_completion_tokens=1024,
+        top_p=0.95,
+        stream=True,
+        reasoning_format="raw",
+    )
+    out = []
+    for chunk in stream:
+        delta = chunk.choices[0].delta.content or ""
+        out.append(delta)
+    return "".join(out)
+
+
+# ---------- UI ----------
 st.set_page_config(page_title="AI Relationship Counsellor", layout="centered")
-st.title("AI Relationship Counsellor")
+st.title("NLP Relationship Counsellor")
 
 uploaded_file = st.file_uploader("Upload a .txt export of your WhatsApp chat", type=["txt"])
 
@@ -128,7 +112,7 @@ if uploaded_file:
         if index.ntotal == 0:
             st.warning("Nothing indexed yet. Please upload a chat file.")
         else:
-            # Retrieve top-k relevant chunks
+            # top-k retrieval
             k = min(5, index.ntotal)
             q_emb = embedding_model.encode([user_query], convert_to_numpy=True).astype("float32")
             distances, idxs = index.search(q_emb, k)
@@ -137,7 +121,7 @@ if uploaded_file:
             context = " ".join(relevant)
             final_prompt = f"Context:\n{context}\n\nQuestion:\n{user_query}"
 
-            with st.spinner("Analyzing… please wait."):
+            with st.spinner("Analyzing…"):
                 answer = query_llm(final_prompt)
 
             st.markdown("### AI Analysis")