Add GPT-4o and GPT-4o-mini integration with NeMo Guardrails

Implemented support for GPT-4o and GPT-4o-mini models, including their integration with NeMo Guardrails, and updated model references across the project files. Also added the new library dependence in requirements.txt and improved code formatting for readability.

app.py

@@ -451,7 +451,9 @@ with gr.Blocks(
 
     with gr.Tab(label="🏆 Leaderboard", elem_id="leaderboard") as leaderboard_tab:
         gr.Markdown("## 🏆 Guardrails Leaderboard")
-        rankings = gr.Markdown("We will launch the guardrails leaderboard once enough votes are collected. Ranking will be calculated based on ELO ratings. Keep playing so that we can collect enough data.")
+        rankings = gr.Markdown(
+            "We will launch the guardrails leaderboard once enough votes are collected. Ranking will be calculated based on ELO ratings. Keep playing so that we can collect enough data."
+        )
         # leaderboard_tab.select(get_rankings, None, [rankings])
 
     gr.Markdown(

guardrails_models.py

@@ -29,7 +29,57 @@ def gpt35_turbo(
     llm = ChatOpenAI(
         temperature=temperature,
         max_retries=6,
-        model_name="gpt-3.5-turbo-1106",
+        model_name="gpt-3.5-turbo",
+        metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
+    )
+    history_langchain_format = []
+    history_langchain_format.append(SystemMessage(system_prompt))
+    for human, ai in history:
+        history_langchain_format.append(HumanMessage(human))
+        if ai:
+            history_langchain_format.append(AIMessage(ai))
+
+    ai_message = llm.stream(history_langchain_format)
+    for message in ai_message:
+        yield message.content
+
+
+def gpt4o(
+    history: List[List[Optional[str]]],
+    system_prompt: str,
+    temperature: float = 1,
+    top_p: float = 0.9,
+    max_output_tokens: int = 2048,
+):
+    llm = ChatOpenAI(
+        temperature=temperature,
+        max_retries=6,
+        model_name="gpt-4o",
+        metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
+    )
+    history_langchain_format = []
+    history_langchain_format.append(SystemMessage(system_prompt))
+    for human, ai in history:
+        history_langchain_format.append(HumanMessage(human))
+        if ai:
+            history_langchain_format.append(AIMessage(ai))
+
+    ai_message = llm.stream(history_langchain_format)
+    for message in ai_message:
+        yield message.content
+
+
+def gpt4o_mini(
+    history: List[List[Optional[str]]],
+    system_prompt: str,
+    temperature: float = 1,
+    top_p: float = 0.9,
+    max_output_tokens: int = 2048,
+):
+    llm = ChatOpenAI(
+        temperature=temperature,
+        max_retries=6,
+        model_name="gpt-4o-mini",
         metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
     )
     history_langchain_format = []
@@ -149,7 +199,7 @@ def gpt35_turbo_llamaguard(
         llm = ChatOpenAI(
             temperature=temperature,
             max_retries=6,
-            model_name="gpt-3.5-turbo-1106",
+            model_name="gpt-3.5-turbo",
             metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
         )
         history_langchain_format = []
@@ -294,7 +344,65 @@ def gpt35_turbo_nemoguardrails(
     rails = LLMRails(
         config,
         llm=ChatOpenAI(
-            model_name="gpt-3.5-turbo-1106",
+            model_name="gpt-3.5-turbo",
+            temperature=temperature,
+            max_retries=6,
+            metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
+        ),
+    )
+    completion = rails.generate(messages=messages)
+    response = completion.get("content", "")
+    for message in response:
+        yield message
+
+
+def gpt4o_nemoguardrails(
+    history: List[List[str]],
+    system_prompt: str,
+    temperature: float = 1,
+    top_p: float = 0.9,
+    max_output_tokens: int = 2048,
+):
+    messages = []
+    messages.append({"role": "system", "content": system_prompt})
+    for human, ai in history:
+        messages.append({"role": "user", "content": human})
+        if ai:
+            messages.append({"role": "assistant", "content": ai})
+    config = RailsConfig.from_path("./nemoguardrails_config")
+    rails = LLMRails(
+        config,
+        llm=ChatOpenAI(
+            model_name="gpt-4o",
+            temperature=temperature,
+            max_retries=6,
+            metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
+        ),
+    )
+    completion = rails.generate(messages=messages)
+    response = completion.get("content", "")
+    for message in response:
+        yield message
+
+
+def gpt4o_mini_nemoguardrails(
+    history: List[List[str]],
+    system_prompt: str,
+    temperature: float = 1,
+    top_p: float = 0.9,
+    max_output_tokens: int = 2048,
+):
+    messages = []
+    messages.append({"role": "system", "content": system_prompt})
+    for human, ai in history:
+        messages.append({"role": "user", "content": human})
+        if ai:
+            messages.append({"role": "assistant", "content": ai})
+    config = RailsConfig.from_path("./nemoguardrails_config")
+    rails = LLMRails(
+        config,
+        llm=ChatOpenAI(
+            model_name="gpt-4o-mini",
             temperature=temperature,
             max_retries=6,
             metadata={"top_p": top_p, "max_output_tokens": max_output_tokens},
@@ -402,23 +510,31 @@ def gemini_pro_nemoguardrails(
 def get_all_models():
     return [
         {
-            "name": "gpt3.5-turbo-1106",
+            "name": "gpt3.5-turbo",
             "model": gpt35_turbo,
         },
         {
-            "name": "Llama-2-70b-chat-hf",
-            "model": llama70B,
+            "name": "gpt4o",
+            "model": gpt4o,
         },
         {
-            "name": "Mixtral-8x7B-Instruct-v0.1",
-            "model": mixtral7x8,
+            "name": "gpt4o-mini",
+            "model": gpt4o_mini,
         },
         {
             "name": "Gemini-Pro",
             "model": gemini_pro,
         },
         # {
-        #     "name": "gpt3.5-turbo-1106 + Llama Guard",
+        #     "name": "Llama-2-70b-chat-hf",
+        #     "model": llama70B,
+        # },
+        # {
+        #     "name": "Mixtral-8x7B-Instruct-v0.1",
+        #     "model": mixtral7x8,
+        # }, "model": gemini_pro,
+        # {
+        #     "name": "gpt3.5-turbo + Llama Guard",
         #     "model": gpt35_turbo_llamaguard,
         # },
         # {
@@ -434,21 +550,29 @@ def get_all_models():
         #     "model": gemini_pro_llamaguard,
         # },
         {
-            "name": "gpt3.5-turbo-1106 + NeMo Guardrails",
+            "name": "gpt3.5-turbo + NeMo Guardrails",
             "model": gpt35_turbo_nemoguardrails,
         },
         {
-            "name": "Llama-2-70b-chat-hf + NeMo Guardrails",
-            "model": llama70B_nemoguardrails,
+            "name": "gpt4o + NeMo Guardrails",
+            "model": gpt4o_nemoguardrails,
         },
         {
-            "name": "Mixtral-8x7B-Instruct-v0.1 + NeMo Guardrails",
-            "model": mixtral7x8_nemoguardrails,
+            "name": "gpt4o-mini + NeMo Guardrails",
+            "model": gpt4o_mini_nemoguardrails,
         },
         {
             "name": "Gemini-Pro + NeMo Guardrails",
             "model": gemini_pro_nemoguardrails,
         },
+        # {
+        #     "name": "Llama-2-70b-chat-hf + NeMo Guardrails",
+        #     "model": llama70B_nemoguardrails,
+        # },
+        # {
+        #     "name": "Mixtral-8x7B-Instruct-v0.1 + NeMo Guardrails",
+        #     "model": mixtral7x8_nemoguardrails,
+        # },
     ]