Show tool calls on the Gradio chat interface.

lynxkite-lynxscribe/src/lynxkite_lynxscribe/agentic.py

@@ -8,6 +8,7 @@ from lynxkite_core import ops
 from lynxscribe.components.task_solver import TaskSolver
 from lynxscribe.components.tool_use import LLM
 from lynxscribe.components.mcp_client import MCPClient
+from lynxscribe.components.rag.rag_chatbot import RAGChatbotResponse
 from lynxscribe.core.llm.base import get_llm_engine
 from lynxscribe.core.models.prompts import Function, Tool
 
@@ -32,12 +33,12 @@ def gradio_chat(agent: dict):
 
     async def respond(message, chat_history):
         await ag.init()
-        response = await ag.llm.ask([*chat_history, {"role": "user", "content": message}])
-        answer = ""
-        async for chunk in response.answer:
-            if chunk.choices and chunk.choices[0].delta.content:
-                answer += chunk.choices[0].delta.content
-            yield answer
+        response = await ag.llm.ask(
+            [*chat_history, {"role": "user", "content": message}],
+            expose_tool_calls=True,
+        )
+        async for messages in rag_chatbot_response_to_gradio(response):
+            yield messages
 
     ag = agent_from_dict(agent, default_model={"name": "openai", "model_name": "gpt-4.1-nano"})
     with gr.Blocks() as demo:
@@ -75,7 +76,7 @@ def agent(
     models = []
     for tool in tools:
         if "model" in tool:
-            models.append(tool["llm"]["model"])
+            models.append(tool["model"])
         elif tool.get("extra_prompt"):
             prompt.append(tool["extra_prompt"])
     params = {
@@ -243,3 +244,42 @@ class Agent:
             ),
         )
         return ask
+
+
+async def rag_chatbot_response_to_gradio(response: RAGChatbotResponse):
+    """The Gradio chatbot interface expects a list of ChatMessage objects to be yielded.
+    We can keep updating the messages to stream the response.
+    """
+    import gradio as gr
+
+    answer: list[gr.ChatMessage] = []
+    yield answer
+    async for chunk in response.answer:
+        if not chunk.choices:
+            continue
+        delta = chunk.choices[0].delta
+        if delta.tool_calls:
+            for tool_call in delta.tool_calls:
+                if not answer or not answer[-1].metadata:
+                    # Create a message for the tool call.
+                    answer.append(
+                        gr.ChatMessage(
+                            role="assistant",
+                            content="",
+                            metadata=dict(
+                                title=f"Using tool: {tool_call.function.name}",
+                                status="pending",
+                            ),
+                        )
+                    )
+        if delta.content is None:
+            pass
+        elif answer and bool(answer[-1].metadata) == bool(delta.role == "tool"):
+            # This chunk is also a tool call (or not), so append to the last message.
+            answer[-1].content += delta.content
+        else:
+            if answer and answer[-1].metadata:
+                # Looks like we're done with the tool.
+                answer[-1].metadata["status"] = "done"
+            answer.append(gr.ChatMessage(role="assistant", content=delta.content))
+        yield answer