diff --git a/ai-hub/integration_tests/conftest.py b/ai-hub/integration_tests/conftest.py
index 11eadcb..a715d04 100644
--- a/ai-hub/integration_tests/conftest.py
+++ b/ai-hub/integration_tests/conftest.py
@@ -43,10 +43,43 @@
     os.environ["SYNC_TEST_USER_ID"] = user_id
     client.headers.update({"X-User-ID": user_id})
 
-    # 2. Add API Providers
-    # For now, it relies on config.yaml having gemini configured by default,
-    # but we can optionally call endpoints if they existed for creating dynamic providers.
-    print("[conftest] Skipping dynamic provider configuration (loaded from config.yaml)...")
+    # 2. Add API Providers and Configure LLM RBAC
+    print("[conftest] Configuring LLM provider and grouping...")
+    
+    # Enable Gemini securely
+    prefs_payload = {
+        "llm": {
+            "active_provider": "gemini",
+            "providers": {
+                "gemini": {
+                    "api_key": os.getenv("GEMINI_API_KEY", ""),
+                    "model": "gemini/gemini-3-flash-preview"
+                }
+            }
+        },
+        "tts": {}, "stt": {}, "statuses": {}
+    }
+    client.put(f"{BASE_URL}/users/me/config", json=prefs_payload)
+    
+    # Establish a Group securely provisioned for AI Usage
+    group_payload = {
+        "name": "Integration Default Group",
+        "description": "Global RBAC group for all integration tasks",
+        "policy": {"llm": ["gemini"]}
+    }
+    r_group = client.post(f"{BASE_URL}/users/admin/groups", json=group_payload)
+    
+    if r_group.status_code == 409:
+        r_groups = client.get(f"{BASE_URL}/users/admin/groups")
+        target_group = next(g for g in r_groups.json() if g["name"] == group_payload["name"])
+        group_id = target_group["id"]
+        # Update policy to ensure it remains clean
+        client.put(f"{BASE_URL}/users/admin/groups/{group_id}", json=group_payload)
+    else:
+        group_id = r_group.json().get("id")
+        
+    # Bind the admin testing account into this fully capable RBAC group
+    client.put(f"{BASE_URL}/users/admin/users/{user_id}/group", json={"group_id": group_id})
 
     # 3. Register Nodes
     print("[conftest] Registering test nodes...")
diff --git a/ai-hub/integration_tests/test_llm_chat.py b/ai-hub/integration_tests/test_llm_chat.py
new file mode 100644
index 0000000..7d583d7
--- /dev/null
+++ b/ai-hub/integration_tests/test_llm_chat.py
@@ -0,0 +1,70 @@
+import os
+import httpx
+import pytest
+import json
+
+BASE_URL = os.getenv("SYNC_TEST_BASE_URL", "http://127.0.0.1:8002/api/v1")
+
+def _headers():
+    return {
+        "X-User-ID": os.environ.get("SYNC_TEST_USER_ID", "")
+    }
+
+def test_create_session_and_chat_gemini():
+    """
+    Test successfully spinning up a fresh session and pushing an LLM prompt
+    over the live SSE connection to verify bidirectional inference health.
+    """
+    user_id = os.environ.get("SYNC_TEST_USER_ID", "")
+    assert user_id, "User ID not found in environment from conftest."
+    
+    with httpx.Client(timeout=30.0) as client:
+        # Step 1: Create a new session bound to Gemini
+        session_payload = {
+            "user_id": user_id,
+            "provider_name": "gemini",
+            "feature_name": "default"
+        }
+        r_sess = client.post(f"{BASE_URL}/sessions/", headers=_headers(), json=session_payload)
+        assert r_sess.status_code == 200, f"Failed to create session: {r_sess.text}"
+        
+        session_data = r_sess.json()
+        session_id = session_data["id"]
+        assert session_id, "Session ID not returned."
+        
+        # Step 2: Send a chat prompt instructing the LLM to echo back a phrase.
+        # This asserts litellm, group RBAC, the context window, and SSE streaming are intact.
+        chat_payload = {
+            "prompt": "I am testing your response capabilities. Please respond with exactly the text 'CORTEX HUB LIVE' and nothing else.",
+            "provider_name": "gemini",
+            "load_faiss_retriever": False
+        }
+        
+        full_response = ""
+        # Endpoint returns an SSE stream, we buffer it by chunks
+        with client.stream("POST", f"{BASE_URL}/sessions/{session_id}/chat", headers=_headers(), json=chat_payload) as r_chat:
+            assert r_chat.status_code == 200, "Chat request failed to initialize."
+            
+            for line in r_chat.iter_lines():
+                if line.startswith("data: "):
+                    data_str = line[len("data: "):]
+                    # Sometimes SSE ends with [DONE]
+                    if data_str == "[DONE]":
+                        break
+                        
+                    try:
+                        event = json.loads(data_str)
+                        event_type = event.get("type")
+                        
+                        if event_type == "content":
+                            full_response += event.get("content", "")
+                        elif event_type == "error":
+                            pytest.fail(f"LLM backend emitted an error: {event.get('content')}")
+                        elif event_type == "done":
+                            break
+                    except json.JSONDecodeError:
+                        pass
+        
+        full_response = full_response.strip()
+        assert len(full_response) > 0, "LLM returned an entirely silent response."
+        assert "CORTEX HUB LIVE" in full_response.upper(), f"LLM missed the instruction formatting. True Response: {full_response}"