feat: RAG integrated with FAISS from HF Dataset + Gemini embedding

app.py

@@ -1,64 +1,141 @@
 import gradio as gr
 import os
-from huggingface_hub import InferenceClient
+import numpy as np
+import pickle
+import requests
+import faiss
+from huggingface_hub import InferenceClient, hf_hub_download
 
-SYSTEM_PROMPT = """You are FeelEd Lite, a friendly Tamil-medium tutor for Grade 11 TN Board Commerce students in Tamil Nadu, India.
-Always answer in simple Tamil + English mix. Keep answers short and clear."""
+# ─── Config ──────────────────────────────
+HF_TOKEN     = os.environ.get("HF_TOKEN", "")
+GEMINI_KEY   = os.environ.get("GEMINI_API_KEY", "")
+DATASET_REPO = "build-small-hackathon/feeled-lite-rag"
 
-HF_TOKEN = os.environ.get("HF_TOKEN", "")
+SYSTEM_PROMPT = """You are FeelEd Lite, a friendly Tamil-medium tutor for TN Board students in Tamil Nadu, India.
+Always answer in simple Tamil + English mix (Tanglish) that a student can easily understand.
+Use the provided textbook context to give accurate answers.
+Keep answers short, clear, and encouraging.
+For Story Mode: explain as a simple story with Tamil characters.
+For Exam Mode: give TN Board exam questions with model answers."""
 
-def generate_response(user_message, mode):
-    if not HF_TOKEN:
-        return "❌ HF_TOKEN not set!", "none"
+# ─── Load FAISS index ─────────────────────
+print("Loading FAISS index from HF Dataset...")
+faiss_index = None
+metadata_store = []
+
+try:
+    index_path = hf_hub_download(
+        repo_id=DATASET_REPO,
+        filename="index.faiss",
+        repo_type="dataset",
+        token=HF_TOKEN if HF_TOKEN else None,
+    )
+    meta_path = hf_hub_download(
+        repo_id=DATASET_REPO,
+        filename="metadata.pkl",
+        repo_type="dataset",
+        token=HF_TOKEN if HF_TOKEN else None,
+    )
+    faiss_index = faiss.read_index(index_path)
+    with open(meta_path, "rb") as f:
+        data = pickle.load(f)
+        metadata_store = data["metadata"]
+    print(f"✅ FAISS loaded: {faiss_index.ntotal} vectors")
+except Exception as e:
+    print(f"⚠️ FAISS load failed: {e}")
+
+# ─── Gemini Embed ─────────────────────────
+def embed_query(text: str):
+    if not GEMINI_KEY:
+        return None
+    url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:embedContent?key={GEMINI_KEY}"
+    body = {"content": {"parts": [{"text": text}]}, "outputDimensionality": 768}
+    try:
+        r = requests.post(url, json=body, timeout=10)
+        if r.ok:
+            return r.json().get("embedding", {}).get("values", [])
+    except:
+        pass
+    return None
+
+# ─── RAG Search ───────────────────────────
+def rag_search(query: str, grade: str = "", subject: str = "", top_k: int = 5) -> str:
+    if faiss_index is None or not GEMINI_KEY:
+        return ""
+    vec = embed_query(query)
+    if not vec:
+        return ""
+    q = np.array([vec], dtype=np.float32)
+    faiss.normalize_L2(q)
+    scores, indices = faiss_index.search(q, top_k * 3)
+    chunks = []
+    for score, idx in zip(scores[0], indices[0]):
+        if idx < 0 or score < 0.35:
+            continue
+        meta = metadata_store[idx]
+        if grade and meta.get("grade") != grade:
+            continue
+        text = meta.get("text", "").strip()
+        if text and len(text) > 30:
+            chunks.append(text)
+        if len(chunks) >= top_k:
+            break
+    return "\n\n".join(chunks[:top_k])
+
+# ─── Inference ────────────────────────────
+def generate_response(user_message, mode, grade="11", subject="Commerce"):
+    context = rag_search(user_message, grade=grade)
 
     if mode == "📚 Q&A":
-        user = f"Grade 11 TN Board Commerce: {user_message}\nAnswer simply in Tamil+English:"
+        user = f"""Textbook Context:
+{context if context else '(No specific context found)'}
+
+Student Question: {user_message}
+Answer simply in Tamil+English mix:"""
     elif mode == "📖 Story Mode":
-        user = f"Explain as a short simple story with Tamil characters: {user_message}"
+        user = f"""Context: {context[:500] if context else ''}
+
+Explain this topic as a short simple story with Tamil characters: {user_message}"""
     else:
-        user = f"3 TN Board exam questions with answers for Grade 11 Commerce: {user_message}"
+        user = f"""Context: {context[:500] if context else ''}
+
+Give 3 important TN Board exam questions with model answers for: {user_message}"""
 
-    # provider + model combinations that work on free tier
     combos = [
         ("Qwen/Qwen2.5-7B-Instruct", "novita"),
         ("meta-llama/Llama-3.1-8B-Instruct", "novita"),
         ("Qwen/Qwen2.5-3B-Instruct", "novita"),
-        ("meta-llama/Llama-3.2-3B-Instruct", "novita"),
         ("Qwen/Qwen2.5-72B-Instruct", "nebius"),
-        ("mistralai/Mistral-7B-Instruct-v0.3", "hf-inference"),
     ]
 
     errors = []
     for model_id, provider in combos:
         try:
-            client = InferenceClient(
-                model=model_id,
-                token=HF_TOKEN,
-                provider=provider,
-            )
+            client = InferenceClient(model=model_id, token=HF_TOKEN, provider=provider)
             response = client.chat_completion(
                 messages=[
                     {"role": "system", "content": SYSTEM_PROMPT},
                     {"role": "user", "content": user}
                 ],
-                max_tokens=350,
+                max_tokens=400,
                 temperature=0.7,
             )
             return response.choices[0].message.content.strip(), model_id.split("/")[-1]
         except Exception as e:
-            errors.append(f"{model_id.split('/')[-1]}({provider}): {str(e)[:120]}")
+            errors.append(f"{model_id.split('/')[-1]}: {str(e)[:80]}")
             continue
 
-    return f"எல்லா models-உம் fail:\n" + "\n".join(errors), "none"
+    return "மன்னிக்கவும், இப்போது busy. சற்று நேரம் கழித்து முயற்சிக்கவும்.", "none"
 
 def chat(message, history, mode):
     if not message.strip():
         return history, "", "🤖 Ready"
     history = history or []
+    rag_status = "📚 RAG: ✅" if faiss_index else "📚 RAG: ❌"
     response, used_model = generate_response(message, mode)
     history.append({"role": "user", "content": message})
     history.append({"role": "assistant", "content": response})
-    return history, "", f"🤖 {used_model}"
+    return history, "", f"🤖 {used_model} | {rag_status}"
 
 CSS = """
 @import url('https://fonts.googleapis.com/css2?family=Noto+Sans+Tamil:wght@400;600;700&family=Inter:wght@400;500;600&display=swap');
@@ -70,13 +147,15 @@ body, .gradio-container { background: #0f0f1a !important; font-family: 'Inter',
 #info-box { background: #16213e; border: 1px solid #2d3748; border-radius: 10px; padding: 14px; color: #718096; font-size: 0.82rem; line-height: 1.8; margin-top: 12px; }
 """
 
+rag_status_text = f"✅ {faiss_index.ntotal} vectors loaded" if faiss_index else "⚠️ RAG not loaded"
+
 with gr.Blocks(css=CSS, title="FeelEd Lite — Tamil TN Tutor") as demo:
-    gr.HTML("""
+    gr.HTML(f"""
     <div id="header">
         <h1>📚 FeelEd Lite</h1>
-        <p>Tamil Medium · Grade 11 Commerce · TN Board Tutor</p>
+        <p>Tamil Medium · Grades 9-12 · TN Board Tutor</p>
         <p>தமிழ் மீடியம் மாணவர்களுக்கான AI கல்வி உதவியாளர்</p>
-        <span class="model-badge">🤖 Small Model · Build Small Hackathon 2026</span>
+        <span class="model-badge">🤖 Small Model · 📚 RAG: {rag_status_text} · Build Small Hackathon 2026</span>
     </div>
     """)
 
@@ -85,24 +164,38 @@ with gr.Blocks(css=CSS, title="FeelEd Lite — Tamil TN Tutor") as demo:
             chatbot = gr.Chatbot(type="messages", height=400, show_label=False, bubble_full_width=False)
             model_display = gr.Textbox(value="🤖 Ready", show_label=False, interactive=False, container=False)
             with gr.Row():
-                msg = gr.Textbox(placeholder="உங்கள் கேள்வியை கேளுங்கள்...", show_label=False, scale=5, lines=2)
+                msg = gr.Textbox(
+                    placeholder="உங்கள் கேள்வியை Tamil அல்லது English-ல கேளுங்கள்...",
+                    show_label=False, scale=5, lines=2,
+                )
                 with gr.Column(scale=1, min_width=90):
                     send_btn = gr.Button("அனுப்பு ▶", variant="primary")
                     clear_btn = gr.Button("🗑 Clear")
 
         with gr.Column(scale=1, min_width=200):
-            mode = gr.Radio(choices=["📚 Q&A", "📖 Story Mode", "🎯 Exam Mode"], value="📚 Q&A", label="📌 Mode:")
+            mode = gr.Radio(
+                choices=["📚 Q&A", "📖 Story Mode", "🎯 Exam Mode"],
+                value="📚 Q&A", label="📌 Mode:",
+            )
+            gr.HTML("<hr style='border-color:#2d3748;margin:12px 0;'>")
             gr.Examples(
-                examples=[["தேவை விதி என்றால் என்ன?"], ["GDP explain"], ["தொழில் முனைவோர் பண்புகள்"], ["Consumer rights"]],
+                examples=[
+                    ["தேவை விதி என்றால் என்ன?"],
+                    ["இரட்டை பதிவு முறை விளக்கு"],
+                    ["GDP என்றால் என்ன?"],
+                    ["தொழில் முனைவோர் பண்புகள்"],
+                    ["Consumer rights explain"],
+                ],
                 inputs=msg, label="💡 Examples:",
             )
-            gr.HTML("""
+            gr.HTML(f"""
             <div id='info-box'>
-                <strong style='color:#63b3ed;'>FeelEd Lite v0.1</strong><br>
-                🏫 Grade 11 TN Commerce<br>
+                <strong style='color:#63b3ed;'>FeelEd Lite v0.2</strong><br>
+                🏫 Grades 9-12 TN Board<br>
+                📚 RAG: TN Textbooks<br>
                 🌐 Tamil + English<br>
                 🔒 Student privacy first<br><br>
-                <em style='color:#4a5568;'>Build Small Hackathon 2026</em>
+                <em style='color:#4a5568;'>Build Small Hackathon 2026<br>Track: Backyard AI</em>
             </div>
             """)
 

requirements.txt

@@ -1,4 +1,4 @@
 huggingface_hub>=0.24.0
 faiss-cpu>=1.7.4
 numpy>=1.24.0
-sentence-transformers>=2.2.0
+requests>=2.28.0