+ robust error handling

agents.py

@@ -1,6 +1,7 @@
 import os
 import random
 import asyncio
+import ssl
 from dotenv import load_dotenv
 from llama_index.core.agent.workflow import AgentWorkflow
 from llama_index.llms.huggingface_api import HuggingFaceInferenceAPI
@@ -13,6 +14,7 @@ from opentelemetry.sdk.trace import TracerProvider
 from langfuse import get_client
 from rich.pretty import pprint
 import aiohttp
+from tenacity import retry, stop_after_attempt, wait_exponential, retry_if_exception_type
 
 
 
@@ -86,6 +88,17 @@ class LlamaIndexAgent:
         )
         
 
+    @retry(
+        stop=stop_after_attempt(3),
+        wait=wait_exponential(multiplier=1, min=4, max=10),
+        retry=retry_if_exception_type((
+            aiohttp.client_exceptions.ClientConnectionError,
+            aiohttp.client_exceptions.ClientOSError,
+            ssl.SSLError,
+            KeyError,
+            ConnectionError
+        ))
+    )
     async def run_query(self, query: str):
         # Instrument LlamaIndex with OpenTelemetry
         self.instrumentor.instrument()
@@ -101,7 +114,16 @@ class LlamaIndexAgent:
                 response = await self.alfred.run(query)
             except aiohttp.client_exceptions.ClientConnectionError as e:
                 span.update_trace(output={"response": f"Connection error: {e}"})
-                return f"AGENT ERROR: Connection error to LLM API: {e}"
+                raise  # Re-raise for retry logic
+            except aiohttp.client_exceptions.ClientOSError as e:
+                span.update_trace(output={"response": f"Client OS error: {e}"})
+                raise  # Re-raise for retry logic
+            except ssl.SSLError as e:
+                span.update_trace(output={"response": f"SSL error: {e}"})
+                raise  # Re-raise for retry logic
+            except (KeyError, ConnectionError) as e:
+                span.update_trace(output={"response": f"Session/Connection error: {e}"})
+                raise  # Re-raise for retry logic
             except Exception as e:
                 span.update_trace(output={"response": f"General error: {e}"})
                 return f"AGENT ERROR: {e}"