Make a Langfuse trace worth opening: question, answer, session, no probe noise

apps/ai-service/main.py

@@ -1,11 +1,13 @@
 from __future__ import annotations
 
+from contextlib import nullcontext
 from hmac import compare_digest
 from time import perf_counter
 from typing import Any
 
 from fastapi import FastAPI, Request, Response
 
+from adapters.langfuse_scores import LangfuseScoreClient
 from adapters.postgres import PostgresTraceRepository
 from bootstrap import build_runtime
 from config import Settings, get_settings
@@ -38,6 +40,10 @@ def create_app(
     app.state.conversational = conversational
     app.state.trace_writer = trace_writer
     app.state.metrics = metrics
+    # None unless Langfuse is fully configured; `/v1/rag/feedback` then skips
+    # mirroring the user's verdict there. Built from the same settings the
+    # span exporter reads, so the two can never point at different projects.
+    app.state.langfuse_scores = LangfuseScoreClient.from_settings(configured)
     # The raw Qdrant retriever, not routed through RagAgent — Feature-List
     # #4/#23's section-list and verbatim-section-text endpoints are plain
     # payload-filtered reads with no LLM/generation step, so nothing about
@@ -51,8 +57,19 @@ def create_app(
         method = request.method.upper()
         started = perf_counter()
         status_code = 500
+        # Kubernetes probes and the Prometheus scrape run every few seconds
+        # forever and carry no information a trace could show. Left traced they
+        # were ~90% of all spans within a day, burying the RAG turns anyone
+        # actually opens a trace viewer to read, and costing storage and
+        # ingestion for nothing. Their Prometheus counters below are untouched,
+        # which is the right place to watch probe health anyway.
+        span_context = (
+            nullcontext(None)
+            if route in _UNTRACED_ROUTES
+            else request_span(method, route, request.headers)
+        )
         with correlation_context(request.headers.get("x-correlation-id")) as correlation_id:
-            with request_span(method, route, request.headers) as span:
+            with span_context as span:
                 try:
                     response = await call_next(request)
                     status_code = response.status_code
@@ -126,6 +143,9 @@ def create_app(
     return app
 
 
+_UNTRACED_ROUTES = frozenset({"/health", "/ready", "/metrics"})
+
+
 def _route_label(path: str) -> str:
     known = {
         "/health", "/ready", "/metrics", "/v1/rag/query", "/v1/rag/suggest",