add modal.py

modal.py

@@ -0,0 +1,89 @@
+from smolagents import CodeAgent, LiteLLMModel
+from smolagents.tools import ToolCollection
+import gradio as gr
+import modal
+import subprocess
+import time
+
+app = modal.App("agent-ontology")
+
+OLLAMA_IMAGE = (
+    modal.Image.debian_slim()
+    # pkill/pgrep come from procps
+    .apt_install(
+        "curl", "gnupg", "software-properties-common",
+        "procps"               # ← adds pkill, pgrep, ps …
+    )
+    # install Ollama
+    .run_commands("curl -fsSL https://ollama.com/install.sh | sh")
+    # spin up daemon, pull the model, shut daemon down
+    .run_commands(
+        "bash -c 'ollama serve >/dev/null 2>&1 & "
+        "PID=$!; "
+        "sleep 10 && "
+        "ollama pull devstral:latest && "
+        "kill $PID'"
+    )
+    # python deps
+    .pip_install(
+        "fastmcp>=2.6.1",
+        "gradio[mcp]>=5.0.0",
+        "huggingface_hub[mcp]>=0.32.2",
+        "mcp>=1.9.2",
+        "smolagents[litellm,mcp]>=1.17.0",
+        "textblob>=0.19.0",
+    )
+)
+
+# Specify the dependencies in the Modal function
+@app.function(image=OLLAMA_IMAGE, gpu="A10G", timeout=2400)
+def run_agent():
+
+    def chat_with_agent(message, history):
+        """Initialize MCP client for each request to avoid connection issues"""
+        try:
+            with ToolCollection.from_mcp(
+                {"url": "https://notredameslab-nf-ontology.hf.space/gradio_api/mcp/sse", "transport": "sse"},
+                trust_remote_code=True  # Acknowledge that we trust this remote MCP server
+            ) as tool_collection:
+                
+                model = LiteLLMModel(
+                    model_id="ollama/devstral:latest",
+                    api_base="http://localhost:11434",
+                )
+
+                agent = CodeAgent(
+                    tools=tool_collection.tools,
+                    model=model,
+                    additional_authorized_imports=["inspect", "json"]
+                )
+
+                additional_instructions = """
+                ADDITIONAL IMPORTANT INSTRUCTIONS:
+                use the tool "final_answer" in the code block to provide the answer to the user. Prints are only for debugging purposes. So, to give your results concatenate everything you want to print in a single "final_answer" call as such : final_answer(f"your answer here").
+                """
+
+                agent.system_prompt += additional_instructions
+
+                result = agent.run(message)
+                return str(result)
+                
+        except Exception as e:
+            return f"❌ Error: {e}\nType: {type(e).__name__}"
+
+    demo = gr.ChatInterface(
+        fn=chat_with_agent,
+        type="messages",
+        examples=["can you extract input/output metadata from fastqc nf-core module ?"],
+        title="Agent with MCP Tools (Per-Request Connection)",
+        description="This version creates a new MCP connection for each request."
+    )
+    demo.launch(share=True)
+
+@app.local_entrypoint()
+def main():
+    """Run the Modal app locally."""
+    run_agent.remote()
+
+if __name__ == "__main__":
+    main()