langfuse generation spans with token usage; switch llm to vllm (qwen2.5-coder-7b on mcrndeb)

api/analyses/drill_down/helpers.py

@@ -43,11 +43,13 @@ async def decide_next(question: str, metric: str, dimensions: list[str],
         history=format_history(slices),
         budget=budget,
     )
-    with lf.span("drill_down.next", as_type="generation", input={"budget": budget}) as span:
+    with lf.span("drill_down.next", input={"budget": budget}) as span:
         await events.publish_current(events.llm_call(
             "drill_down.next", system_len=len(system), user_len=len(user),
         ))
-        decision = _parse_json(chat(system=system, user=user, max_tokens=256))
+        decision = _parse_json(chat(
+            system=system, user=user, max_tokens=256, span_name="drill_down.next.gen",
+        ))
 
         # Hard guard: the chosen dimension MUST be in the candidate list.
         if decision.get("action") == "drill":
@@ -142,7 +144,7 @@ async def interpret(question: str, metric: str, slices: list[Slice]) -> str:
     await events.publish_current(events.llm_call(
         "drill_down.interpret", system_len=len(system), user_len=len(user),
     ))
-    return chat(system=system, user=user, max_tokens=512)
+    return chat(system=system, user=user, max_tokens=512, span_name="drill_down.interpret")
 
 
 # ── Prompt-context formatters ──