Handle unset ttft in metadata of failed chat response. Fixes evals.

This was causing evals to stop processing rest of batch as well.

src/khoj/routers/api_chat.py

@@ -809,11 +809,11 @@ async def chat(
             chat_metadata = chat_metadata or {}
             chat_metadata["conversation_command"] = cmd_set
             chat_metadata["agent"] = conversation.agent.slug if conversation and conversation.agent else None
-            chat_metadata["latency"] = f"{latency:.3f}"
-            chat_metadata["ttft_latency"] = f"{ttft:.3f}"
             chat_metadata["cost"] = f"{cost:.5f}"
-
-            logger.info(f"Chat response time to first token: {ttft:.3f} seconds")
+            chat_metadata["latency"] = f"{latency:.3f}"
+            if ttft:
+                chat_metadata["ttft_latency"] = f"{ttft:.3f}"
+                logger.info(f"Chat response time to first token: {ttft:.3f} seconds")
             logger.info(f"Chat response total time: {latency:.3f} seconds")
             logger.info(f"Chat response cost: ${cost:.5f}")
             update_telemetry_state(

tests/evals/eval.py

@@ -516,6 +516,7 @@ def process_batch(batch, batch_start, results, dataset_length, response_evaluato
         if is_none_or_empty(agent_response):
             decision = None
             explanation = "Agent response is empty. This maybe due to a service error."
+            eval_cost = 0.0
         else:
             decision, explanation, eval_cost = response_evaluator(prompt, agent_response, answer, agent_references)