Evaluation: Incomple Langfuse traces sycned

backend/app/crud/evaluations/langfuse.py

@@ -492,28 +492,9 @@ def fetch_trace_scores_from_langfuse(
                 )
                 continue
 
-        # 4. Identify complete scores (all traces must have the score)
-        total_traces = len(traces)
-        complete_score_names = {
-            name
-            for name, data in score_aggregations.items()
-            if len(data["values"]) == total_traces
-        }
-
-        # 5. Filter trace scores to only include complete scores
-        for trace in traces:
-            trace["scores"] = [
-                score
-                for score in trace["scores"]
-                if score["name"] in complete_score_names
-            ]
-
-        # 6. Calculate summary scores (only for complete scores)
+        # 4. Calculate summary scores for all scores that have at least one value
         summary_scores = []
         for score_name, agg_data in score_aggregations.items():
-            if score_name not in complete_score_names:
-                continue
-
             data_type = agg_data["data_type"]
             values = agg_data["values"]
 
@@ -552,7 +533,7 @@ def fetch_trace_scores_from_langfuse(
 
         logger.info(
             f"[fetch_trace_scores_from_langfuse] Successfully fetched scores | "
-            f"total_traces={len(traces)} | complete_scores={list(complete_score_names)}"
+            f"total_traces={len(traces)} | score_names={list(score_aggregations.keys())}"
         )
 
         return result

backend/app/services/evaluations/evaluation.py

@@ -239,7 +239,10 @@ def get_evaluation_with_scores(
     if not get_trace_info:
         return eval_run, None
 
-    # Check if we already have cached traces
+    # Caching strategy: On the first request, trace scores are fetched from Langfuse
+    # and a copy is stored in S3. Subsequent requests serve from
+    # S3 instead of Langfuse, which is significantly faster. Use resync_score=true
+    # to bypass the cache and re-fetch from Langfuse and store to S3 again.
     has_cached_traces_s3 = eval_run.score_trace_url is not None
     has_cached_traces_db = eval_run.score is not None and "traces" in eval_run.score
     if not resync_score:
@@ -276,6 +279,7 @@ def get_evaluation_with_scores(
             )
             return eval_run, None
 
+    # Resync requested — fetch fresh scores from Langfuse
     langfuse = get_langfuse_client(
         session=session,
         org_id=organization_id,
@@ -321,7 +325,7 @@ def get_evaluation_with_scores(
 
     merged_summary_scores = list(existing_scores_map.values())
 
-    # Build final score with merged summary_scores and traces
+    # Build final score and persist to S3/DB for future cached reads
     score = {
         "summary_scores": merged_summary_scores,
         "traces": langfuse_score.get("traces", []),