Verify exact production traces and record rollout

2026-08-11 11:29:59 +07:00
parent 6b8f7584ed
commit 59e6ad2d0d
46 changed files with 2795 additions and 290 deletions
@@ -26,6 +26,9 @@ RUN pip install --no-cache-dir \
    "qdrant-client>=1.7,<2" \
    "uvicorn[standard]>=0.30,<1" \
    "prometheus-client>=0.20,<1" \
+    "opentelemetry-api>=1.27,<2" \
+    "opentelemetry-sdk>=1.27,<2" \
+    "opentelemetry-exporter-otlp-proto-http>=1.27,<2" \
    "anthropic>=0.112,<1" \
    "boto3"

@@ -18,6 +18,8 @@ class RetrievalTrace:
    reason: str
    resolved_drug_id: str | None
    citations: tuple[dict[str, Any], ...]
+    correlation_id: str | None = None
+    otel_trace_id: str | None = None
    created_at: datetime | None = None


@@ -55,6 +57,8 @@ class PostgresTraceRepository:
        reason: str,
        resolved_drug_id: str | None,
        citations: tuple[dict[str, Any], ...],
+        correlation_id: str | None = None,
+        otel_trace_id: str | None = None,
    ) -> str:
        import psycopg

@@ -64,12 +68,14 @@ class PostgresTraceRepository:
                """
                INSERT INTO rag_retrieval_trace (
                    trace_id, query_text, subject_scope, query_intent,
-                    decision, reason, resolved_drug_id, citations
-                ) VALUES (%s, %s, %s, %s, %s, %s, %s, %s::jsonb)
+                    decision, reason, resolved_drug_id, citations,
+                    correlation_id, otel_trace_id
+                ) VALUES (%s, %s, %s, %s, %s, %s, %s, %s::jsonb, %s, %s)
                """,
                (
                    trace_id, query, subject_scope, intent, decision, reason,
                    resolved_drug_id, json.dumps(citations, ensure_ascii=False),
+                    correlation_id, otel_trace_id,
                ),
            )
        return trace_id
@@ -81,7 +87,8 @@ class PostgresTraceRepository:
            row = connection.execute(
                """
                SELECT trace_id::text, query_text, subject_scope, query_intent,
-                       decision, reason, resolved_drug_id, citations, created_at
+                       decision, reason, resolved_drug_id, citations,
+                       correlation_id, otel_trace_id, created_at
                FROM rag_retrieval_trace WHERE trace_id = %s
                """,
                (trace_id,),
@@ -91,7 +98,8 @@ class PostgresTraceRepository:
        return RetrievalTrace(
            trace_id=row[0], query=row[1], subject_scope=row[2], intent=row[3],
            decision=row[4], reason=row[5], resolved_drug_id=row[6],
-            citations=tuple(row[7]), created_at=row[8],
+            citations=tuple(row[7]), correlation_id=row[8], otel_trace_id=row[9],
+            created_at=row[10],
        )


@@ -1,11 +1,9 @@
-"""Exports the domain's counters; the only module that names prometheus_client.
+"""Prometheus export for bounded domain and request-path telemetry.

-`rag/metrics.py` defines what is counted and why. This decides how it leaves
-the process, and is imported lazily so the service runs — and the suite passes
-— with no metrics stack installed.
-
-Counter names carry a `duocthu_` prefix and a `_total` suffix because that is
-what Prometheus expects of a counter; the dashboard queries them by name.
+The domain owns metric names in :mod:`rag.metrics`; this adapter owns label
+vocabularies, buckets and OpenMetrics exposition. Unknown label values are
+collapsed to ``other`` so a query, drug id, exception message or raw URL can
+never accidentally create an unbounded time series.
 """
 from __future__ import annotations

@@ -14,44 +12,192 @@ from typing import Any
 from rag.metrics import (
    ABSTENTION,
    ANSWER_EXTRACTIVE,
+    CLARIFY_ASKED,
+    DECISION,
+    FOLLOWUP_INHERITED,
    GENERATION_REJECTED,
    GENERATION_SERVED,
+    LOOP_REFINED,
+    LOOP_REPAIRED,
+    LOOP_ROUNDS,
+    PROVIDER_FAILURE,
+    REQUEST_DURATION,
+    REQUESTS,
    RETRIEVAL_ROUTE,
+    STAGE_DURATION,
+    TRACE_WRITE_FAILED,
 )

-_LABELS: dict[str, tuple[str, ...]] = {
+_COUNTER_LABELS: dict[str, tuple[str, ...]] = {
    ABSTENTION: ("reason",),
-    GENERATION_REJECTED: ("reason",),
-    RETRIEVAL_ROUTE: ("route",),
-    GENERATION_SERVED: (),
    ANSWER_EXTRACTIVE: (),
+    CLARIFY_ASKED: ("reason",),
+    DECISION: ("decision", "reason"),
+    FOLLOWUP_INHERITED: (),
+    GENERATION_REJECTED: ("reason",),
+    GENERATION_SERVED: (),
+    LOOP_REFINED: (),
+    LOOP_REPAIRED: (),
+    LOOP_ROUNDS: (),
+    PROVIDER_FAILURE: ("provider", "operation", "reason"),
+    REQUESTS: ("method", "route", "status"),
+    RETRIEVAL_ROUTE: ("route",),
+    TRACE_WRITE_FAILED: (),
+}
+
+_HISTOGRAM_LABELS: dict[str, tuple[str, ...]] = {
+    REQUEST_DURATION: ("method", "route", "status"),
+    STAGE_DURATION: ("stage", "outcome"),
 }

 _HELP = {
-    ABSTENTION: "Answers refused, by the reason retrieval gave.",
-    GENERATION_REJECTED: (
-        "Generations discarded before reaching the caller. `reason=\"ungrounded_number\"` "
-        "counts answers that stated a figure absent from the cited source; "
-        "`reason=\"uncited_claim\"` counts claims with no valid citation at all; "
-        "`reason=\"unsupported_claim\"` counts claims the entailment pass judged "
-        "not actually stated by the block they cite."
-    ),
-    GENERATION_SERVED: "Generations that passed grounding verification and were served.",
+    ABSTENTION: "Answers refused, by bounded domain reason.",
    ANSWER_EXTRACTIVE: "Answers served as verbatim source text.",
-    RETRIEVAL_ROUTE: "Retrievals by route: section filter, or similarity fallback.",
+    CLARIFY_ASKED: "Clarifying questions returned instead of guessing.",
+    DECISION: "Final request decisions by bounded domain reason.",
+    FOLLOWUP_INHERITED: "Follow-up turns that inherited prior context.",
+    GENERATION_REJECTED: "Generated answers discarded by a safety or availability gate.",
+    GENERATION_SERVED: "Generated answers that passed grounding and entailment.",
+    LOOP_REFINED: "Conversational retrieval loops that refined a query.",
+    LOOP_REPAIRED: "Conversational retrieval loops that repaired an answer.",
+    LOOP_ROUNDS: "Retrieval loop rounds completed.",
+    PROVIDER_FAILURE: "Failures from bounded external-provider operations.",
+    REQUESTS: "HTTP requests by route template and status class.",
+    REQUEST_DURATION: "End-to-end HTTP request latency in seconds.",
+    RETRIEVAL_ROUTE: "Retrievals by bounded route.",
+    STAGE_DURATION: "RAG stage latency in seconds.",
+    TRACE_WRITE_FAILED: "Final trace rows that could not be persisted to PostgreSQL.",
 }

+_REQUEST_BUCKETS = (0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10, 20, 40, 60)
+_STAGE_BUCKETS = (0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10, 20, 40)
+
+_ALLOWED: dict[str, frozenset[str]] = {
+    "decision": frozenset({"answerable", "clarify", "verify_pdf", "abstain", "error"}),
+    "method": frozenset({"GET", "POST", "PUT", "PATCH", "DELETE", "OPTIONS"}),
+    "operation": frozenset(
+        {"understand", "embed", "retrieve", "rerank", "generate", "entailment", "persist"}
+    ),
+    "outcome": frozenset({"ok", "error", "cancelled"}),
+    "provider": frozenset(
+        {"bedrock_converse", "bedrock_claude", "bedrock_cohere", "qdrant", "postgres"}
+    ),
+    "route": frozenset(
+        {
+            "/health",
+            "/ready",
+            "/metrics",
+            "/v1/rag/query",
+            "/v1/rag/suggest",
+            "section",
+            "overview",
+            "similarity",
+            "indication",
+            "other",
+        }
+    ),
+    "stage": frozenset(
+        {
+            "receive",
+            "context",
+            "understanding",
+            "routing",
+            "retrieval",
+            "rerank",
+            "evidence",
+            "generation",
+            "grounding",
+            "entailment",
+            "persistence",
+            "response",
+        }
+    ),
+    "status": frozenset({"1xx", "2xx", "3xx", "4xx", "5xx"}),
+}
+
+_REASONS = frozenset(
+    {
+        "clarify_loop_exhausted",
+        "drug_not_in_formulary",
+        "drug_not_resolved",
+        "drug_resolution_ambiguous",
+        "drug_resolution_invalid_state",
+        "evidence_insufficient",
+        "error",
+        "generation_unavailable",
+        "grounded_evidence_available",
+        "incomplete_answer",
+        "insufficient_retrieval_score",
+        "invalid_citation",
+        "malformed_output",
+        "missing_attribute",
+        "missing_indication",
+        "missing_pediatric_age_or_weight",
+        "missing_population",
+        "missing_printed_page_provenance",
+        "missing_provenance",
+        "missing_query_or_drug",
+        "needs_more_info",
+        "no_drug",
+        "no_indication",
+        "no_indication_match",
+        "no_interaction_evidence",
+        "out_of_scope",
+        "out_of_scope_non_human",
+        "parent_hydration_failed",
+        "provider_unavailable",
+        "query_embedding_unavailable",
+        "query_intent_unknown",
+        "request_budget_exhausted",
+        "smalltalk",
+        "subject_scope_unknown",
+        "timeout",
+        "uncited_claim",
+        "ungrounded_number",
+        "unsupported_claim",
+        "visual_verification_required",
+    }
+)
+
+
+def _bounded(label: str, value: str) -> str:
+    allowed = _REASONS if label == "reason" else _ALLOWED.get(label)
+    if allowed is None:
+        return value
+    return value if value in allowed else "other"
+
+
+def _exemplar() -> dict[str, str] | None:
+    # Lazy import avoids making OpenTelemetry a hard requirement for metrics.
+    try:
+        from rag.telemetry import current_trace_id
+
+        trace_id = current_trace_id()
+    except ImportError:
+        trace_id = None
+    return {"trace_id": trace_id} if trace_id else None
+

 class PrometheusMetrics:
-    """Domain `Metrics` backed by a Prometheus registry."""
+    """Domain ``Metrics`` backed by an isolated Prometheus registry."""

    def __init__(self, registry: Any | None = None) -> None:
-        from prometheus_client import CollectorRegistry, Counter
+        from prometheus_client import CollectorRegistry, Counter, Histogram

        self._registry = registry or CollectorRegistry()
        self._counters = {
            name: Counter(name, _HELP[name], labels, registry=self._registry)
-            for name, labels in _LABELS.items()
+            for name, labels in _COUNTER_LABELS.items()
+        }
+        self._histograms = {
+            name: Histogram(
+                name,
+                _HELP[name],
+                labels,
+                buckets=_REQUEST_BUCKETS if name == REQUEST_DURATION else _STAGE_BUCKETS,
+                registry=self._registry,
+            )
+            for name, labels in _HISTOGRAM_LABELS.items()
        }

    @property
@@ -60,17 +206,28 @@ class PrometheusMetrics:

    def increment(self, name: str, **labels: str) -> None:
        counter = self._counters.get(name)
-        if counter is None:
+        expected = _COUNTER_LABELS.get(name)
+        if counter is None or expected is None or set(labels) != set(expected):
            return
-        # An unexpected label would raise at scrape time, far from its cause.
-        # Metrics must not be able to break a clinical answer, so a mismatch
-        # drops the sample rather than the request.
-        expected = set(_LABELS[name])
-        if set(labels) != expected:
+        safe = {key: _bounded(key, labels[key]) for key in expected}
+        child = counter.labels(**safe) if safe else counter
+        child.inc(exemplar=_exemplar())
+
+    def observe(self, name: str, value: float, **labels: str) -> None:
+        histogram = self._histograms.get(name)
+        expected = _HISTOGRAM_LABELS.get(name)
+        if histogram is None or expected is None or set(labels) != set(expected):
            return
-        (counter.labels(**labels) if labels else counter).inc()
+        safe = {key: _bounded(key, labels[key]) for key in expected}
+        child = histogram.labels(**safe) if safe else histogram
+        child.observe(value, exemplar=_exemplar())

    def render(self) -> tuple[bytes, str]:
-        from prometheus_client import CONTENT_TYPE_LATEST, generate_latest
+        # OpenMetrics exposition preserves histogram exemplars. Grafana uses
+        # their trace_id label to jump from a slow aggregate bucket to Tempo.
+        from prometheus_client.openmetrics.exposition import (
+            CONTENT_TYPE_LATEST,
+            generate_latest,
+        )

        return generate_latest(self._registry), CONTENT_TYPE_LATEST
@@ -6,14 +6,22 @@ from adapters.embedding import BedrockCohereQueryEmbedder
 from adapters.postgres import PostgresConversationStore, PostgresTraceRepository
 from adapters.qdrant import QdrantParentStore, QdrantRetriever
 from config import Settings
-from rag.agent import RagAgent
-from rag.answer import GroundedAnswerService
 from rag.artifacts import load_aliases
 from rag.manifest import MANIFEST_POINT_ID, check_manifest, manifest_collection
+from rag.instrumentation import (
+    InstrumentedEmbedder,
+    InstrumentedGenerator,
+    InstrumentedGroundedAnswerService,
+    InstrumentedQueryUnderstander,
+    InstrumentedRagAgent,
+    InstrumentedReranker,
+    InstrumentedRetrievalService,
+)
 from rag.metrics import NullMetrics
 from rag.routing import CatalogDrugResolver, QueryRoutingService
 from rag.sections import SectionResolver
-from rag.service import EvidencePolicy, RetrievalService
+from rag.service import EvidencePolicy
+from rag.telemetry import configure_telemetry
 from rag.understanding import LlmQueryUnderstander

 # How many aliases to show per candidate drug (F-04 bounds *which* drugs are
@@ -126,6 +134,8 @@ def _verify_corpus_manifest(client, collection: str, embedder, settings: Setting

 def build_runtime(settings: Settings):
    metrics = _build_metrics(settings)
+    effective_metrics = metrics or NullMetrics()
+    configure_telemetry(settings, effective_metrics)
    if settings.embedding_provider == "disabled":
        return None, None, PostgresTraceRepository(settings.postgres_dsn), metrics
    if settings.embedding_provider != "cohere-v4":
@@ -140,8 +150,11 @@ def build_runtime(settings: Settings):
        api_key=settings.qdrant_api_key,
        timeout=30,
    )
-    embedder = BedrockCohereQueryEmbedder(
-        settings.embedding_dimensions, region=settings.aws_region
+    embedder = InstrumentedEmbedder(
+        BedrockCohereQueryEmbedder(
+            settings.embedding_dimensions, region=settings.aws_region
+        ),
+        effective_metrics,
    )
    # F-05: a collection built with one model and queried with another
    # returns hits and raises nothing — the results are just meaningless,
@@ -155,17 +168,25 @@ def build_runtime(settings: Settings):
    # `.resolve()` is no longer on the live query path; `RagAgent` resolves
    # drug identity through `LlmQueryUnderstander` against the same catalog.
    resolver = CatalogDrugResolver(aliases)
-    retrieval = RetrievalService(
+    reranker = _build_reranker(settings)
+    retrieval = InstrumentedRetrievalService(
        QdrantRetriever(client, settings.qdrant_collection, embedder),
        QdrantParentStore(client, settings.qdrant_collection),
        EvidencePolicy(minimum_score=settings.evidence_minimum_score),
        section_resolver=section_resolver,
-        reranker=_build_reranker(settings),
+        reranker=(
+            InstrumentedReranker(reranker, effective_metrics)
+            if reranker is not None
+            else None
+        ),
+        metrics=effective_metrics,
    )
    routing = QueryRoutingService(retrieval, resolver)
    generator = _build_generator(settings)
-    answers = GroundedAnswerService(
-        routing, generator=generator, metrics=metrics or NullMetrics()
+    if generator is not None:
+        generator = InstrumentedGenerator(generator, effective_metrics)
+    answers = InstrumentedGroundedAnswerService(
+        routing, generator=generator, metrics=effective_metrics
    )
    trace_writer = PostgresTraceRepository(settings.postgres_dsn)
    if generator is None:
@@ -175,13 +196,16 @@ def build_runtime(settings: Settings):
        # capability to offer. Answer-only (retrieval-verified, no
        # generation) still works through `answers` directly.
        return answers, None, trace_writer, metrics
-    agent = RagAgent(
-        understander=LlmQueryUnderstander(generator, _catalog_names(aliases), resolver),
+    agent = InstrumentedRagAgent(
+        understander=InstrumentedQueryUnderstander(
+            LlmQueryUnderstander(generator, _catalog_names(aliases), resolver)
+        ),
        retrieval=retrieval,
        answers=answers,
        autocomplete=resolver,
        max_wall_clock_ms=settings.max_wall_clock_ms,
        max_llm_calls_per_turn=settings.max_llm_calls_per_turn,
        store=PostgresConversationStore(settings.postgres_dsn),
+        metrics=effective_metrics,
    )
    return answers, agent, trace_writer, metrics
@@ -25,6 +25,7 @@ class Settings(BaseSettings):
    model_config = SettingsConfigDict(env_file=".env", extra="ignore")

    app_name: str = "vsf-duoc-thu-ai-service"
+    environment: str = "local"
    qdrant_url: str = "http://localhost:6333"
    qdrant_collection: str = "duocthu_v1"
    qdrant_api_key: str | None = None
@@ -50,6 +51,13 @@ class Settings(BaseSettings):
    # never uses it.
    rerank_enabled: bool = False
    metrics_enabled: bool = True
+    # OpenTelemetry is opt-in so the existing EC2 Compose deployment keeps
+    # answering when no collector is present. Docker/Kubernetes observability
+    # profiles enable it and point OTLP/HTTP at their local collector Service.
+    otel_enabled: bool = False
+    otel_service_name: str = "ai-service"
+    otel_exporter_otlp_endpoint: str = "http://localhost:4318/v1/traces"
+    otel_sample_ratio: float = Field(default=1.0, ge=0.0, le=1.0)
    entities_path: Path = _default_entities_path()
    # F-08: a per-turn budget across RagAgent's sequential Bedrock calls
    # (understand, generate, one entailment check on the live agent path).
@@ -1,13 +1,22 @@
 from __future__ import annotations

+from time import perf_counter
 from typing import Any

-from fastapi import FastAPI, Response
+from fastapi import FastAPI, Request, Response

 from adapters.postgres import PostgresTraceRepository
 from bootstrap import build_runtime
 from config import Settings, get_settings
 from rag.answer import GroundedAnswerService
+from rag.metrics import REQUEST_DURATION, REQUESTS, NullMetrics
+from rag.telemetry import (
+    annotate_current_span,
+    configure_telemetry,
+    correlation_context,
+    current_trace_id,
+    request_span,
+)
 from routers.rag import router as rag_router


@@ -20,16 +29,65 @@ def create_app(
    metrics: Any | None = None,
 ) -> FastAPI:
    configured = settings or get_settings()
+    effective_metrics = metrics or NullMetrics()
+    configure_telemetry(configured, effective_metrics)
    app = FastAPI(title=configured.app_name, version="0.1.0")
    app.state.answer_service = answer_service
    app.state.conversational = conversational
    app.state.trace_writer = trace_writer
    app.state.metrics = metrics

+    @app.middleware("http")
+    async def correlate_and_trace(request: Request, call_next):
+        route = _route_label(request.url.path)
+        method = request.method.upper()
+        started = perf_counter()
+        status_code = 500
+        with correlation_context(request.headers.get("x-correlation-id")) as correlation_id:
+            with request_span(method, route, request.headers) as span:
+                try:
+                    response = await call_next(request)
+                    status_code = response.status_code
+                    response.headers["X-Correlation-ID"] = correlation_id
+                    trace_id = current_trace_id()
+                    if trace_id:
+                        response.headers["X-Trace-ID"] = trace_id
+                    return response
+                finally:
+                    status = f"{status_code // 100}xx"
+                    elapsed = perf_counter() - started
+                    effective_metrics.increment(
+                        REQUESTS, method=method, route=route, status=status
+                    )
+                    effective_metrics.observe(
+                        REQUEST_DURATION,
+                        elapsed,
+                        method=method,
+                        route=route,
+                        status=status,
+                    )
+                    if span is not None:
+                        span.set_attribute("http.response.status_code", status_code)
+                    annotate_current_span(
+                        **{
+                            "duocthu.http.status_class": status,
+                            "duocthu.duration_ms": elapsed * 1000,
+                        }
+                    )
+
    @app.get("/health")
    def health() -> dict[str, str]:
        return {"status": "ok"}

+    @app.get("/ready")
+    def ready() -> Response:
+        # Runtime construction already verifies the Qdrant corpus manifest and
+        # fails startup on mismatch. PostgreSQL trace/history are intentionally
+        # fail-open and therefore must not make readiness flap.
+        if app.state.answer_service is None and configured.embedding_provider != "disabled":
+            return Response(status_code=503)
+        return Response(content='{"status":"ready"}', media_type="application/json")
+
    @app.get("/metrics")
    def prometheus_metrics() -> Response:
        exporter = getattr(app.state, "metrics", None)
@@ -44,6 +102,11 @@ def create_app(
    return app


+def _route_label(path: str) -> str:
+    known = {"/health", "/ready", "/metrics", "/v1/rag/query", "/v1/rag/suggest"}
+    return path if path in known else "other"
+
+
 _settings = get_settings()
 _answer_service, _conversational, _trace_writer, _metrics = build_runtime(_settings)
 app = create_app(
@@ -1,6 +1,6 @@
 from pathlib import Path

-from adapters.postgres import PostgresConversationStore, PostgresTraceRepository
+from adapters.postgres import PostgresTraceRepository
 from config import get_settings


@@ -8,13 +8,10 @@ def main() -> None:
    dsn = get_settings().postgres_dsn
    migrations_dir = Path(__file__).parent / "migrations"

-    trace_migration = migrations_dir / "001_rag_retrieval_trace.sql"
-    PostgresTraceRepository(dsn).migrate(trace_migration)
-    print(f"Applied {trace_migration.name}")
-
-    conversation_migration = migrations_dir / "002_rag_conversation_turn.sql"
-    PostgresConversationStore(dsn).migrate(conversation_migration)
-    print(f"Applied {conversation_migration.name}")
+    repository = PostgresTraceRepository(dsn)
+    for migration in sorted(migrations_dir.glob("*.sql")):
+        repository.migrate(migration)
+        print(f"Applied {migration.name}")


 if __name__ == "__main__":
@@ -0,0 +1,11 @@
+ALTER TABLE rag_retrieval_trace
+    ADD COLUMN IF NOT EXISTS correlation_id text,
+    ADD COLUMN IF NOT EXISTS otel_trace_id varchar(32);
+
+CREATE INDEX IF NOT EXISTS rag_retrieval_trace_correlation_idx
+    ON rag_retrieval_trace (correlation_id)
+    WHERE correlation_id IS NOT NULL;
+
+CREATE INDEX IF NOT EXISTS rag_retrieval_trace_otel_idx
+    ON rag_retrieval_trace (otel_trace_id)
+    WHERE otel_trace_id IS NOT NULL;
@@ -18,6 +18,11 @@ test = ["pytest>=7.4,<9"]
 # without a cloud generator. Neither is a precondition for a grounded answer.
 metrics = ["prometheus-client>=0.20,<1"]
 generation = ["anthropic>=0.112,<1"]
+observability = [
+    "opentelemetry-api>=1.27,<2",
+    "opentelemetry-sdk>=1.27,<2",
+    "opentelemetry-exporter-otlp-proto-http>=1.27,<2",
+]

 [build-system]
 requires = ["setuptools>=68"]
@@ -99,12 +99,30 @@ class AnswerPlan:
    needs_warning: bool = False


+# A fixed, non-LLM string. `docs/architecture.md`'s guardrail section
+# specifies the disclaimer at several layers, and the web banner
+# (`packages/ui/src/DisclaimerBanner.tsx`) was the only one in place: the
+# `disclaimer` field declared in `packages/shared-types/src/dto/chat.ts` was
+# never populated, so any consumer other than this one web UI received medical
+# content with nothing attached. Keeping it out of the prompt is deliberate —
+# a disclaimer the model writes is one the model can also reword, shorten or
+# omit, and it would then have to be verified like any other generated claim.
+DISCLAIMER = (
+    "Nội dung được trích từ Dược thư Quốc gia Việt Nam 2018, phục vụ tra cứu "
+    "chuyên môn và không thay thế chỉ định của bác sĩ hoặc dược sĩ lâm sàng."
+)
+
+
@dataclass(frozen=True)
 class GroundedAnswer:
    result: RetrievalResult
    answer: str | None
    citations: tuple[Citation, ...] = ()
    generated: bool = False
+    # Carried on the payload rather than added by the caller, so no response
+    # path can be built that omits it — including abstains and clarifications,
+    # which are also medical content in the sense that matters here.
+    disclaimer: str = DISCLAIMER
    # Set when the model decided the turn is under-specified and asked back
    # (e.g. a dose question with no age/weight). The answer field carries the
    # question; the caller renders it as a clarification, not a final answer.
@@ -41,6 +41,8 @@ from __future__ import annotations
 import re
 from dataclasses import dataclass

+from .telemetry import traced_stage
+
 # A digit run with internal separators kept: "500", "7,5", "1.000".
 # Ranges ("4 - 6 giờ") yield two tokens, and each is checked on its own.
 _NUMBER = re.compile(r"\d+(?:[.,]\d+)*")
@@ -127,6 +129,7 @@ def split_claims(answer: str, evidence_count: int) -> tuple[Claim, ...]:
    return tuple(claims)


+@traced_stage("grounding")
 def verify(answer: str, evidence_texts: tuple[str, ...]) -> GroundingReport:
    """Whether `answer` states only figures and sources traceable to the
    specific evidence block(s) cited immediately after each claim.
@@ -0,0 +1,240 @@
+"""Non-invasive instrumentation wrappers for the live RAG object graph.
+
+Claude owns ``rag/agent.py`` and ``rag/answer.py`` in the shared worktree.
+Subclasses here add spans at their stable method boundaries without changing
+those files or duplicating their domain decisions.
+"""
+from __future__ import annotations
+
+from typing import Any
+
+from .agent import RagAgent
+from .answer import GroundedAnswerService
+from .metrics import CLARIFY_ASKED, Metrics, PROVIDER_FAILURE, RETRIEVAL_ROUTE
+from .service import RetrievalService
+from .telemetry import (
+    annotate_current_span,
+    current_stage,
+    dependency_span,
+    stage,
+)
+
+
+def _provider_name(delegate: Any) -> str:
+    name = delegate.__class__.__name__.casefold()
+    if "cohere" in name:
+        return "bedrock_cohere"
+    if "converse" in name:
+        return "bedrock_converse"
+    if "claude" in name:
+        return "bedrock_claude"
+    return "other"
+
+
+def _failure_reason(exc: BaseException) -> str:
+    name = exc.__class__.__name__.casefold()
+    if "timeout" in name:
+        return "timeout"
+    if "budget" in name:
+        return "request_budget_exhausted"
+    if "unavailable" in name or "connection" in name:
+        return "provider_unavailable"
+    return "error"
+
+
+class InstrumentedGenerator:
+    def __init__(self, delegate: Any, metrics: Metrics) -> None:
+        self._delegate = delegate
+        self._metrics = metrics
+        self._provider = _provider_name(delegate)
+
+    @property
+    def model_id(self) -> str:
+        return self._delegate.model_id
+
+    def generate(self, system: str, user: str, schema: dict) -> str:
+        operation = {
+            "understanding": "understand",
+            "generation": "generate",
+            "entailment": "entailment",
+        }.get(current_stage(), "generate")
+        with dependency_span(self._provider, operation):
+            try:
+                return self._delegate.generate(system, user, schema)
+            except Exception as exc:
+                self._metrics.increment(
+                    PROVIDER_FAILURE,
+                    provider=self._provider,
+                    operation=operation,
+                    reason=_failure_reason(exc),
+                )
+                raise
+
+
+class InstrumentedEmbedder:
+    def __init__(self, delegate: Any, metrics: Metrics) -> None:
+        self._delegate = delegate
+        self._metrics = metrics
+
+    @property
+    def dimensions(self) -> int:
+        return self._delegate.dimensions
+
+    @property
+    def model_id(self) -> str:
+        return self._delegate.model_id
+
+    def embed_query(self, text: str) -> list[float]:
+        with dependency_span("bedrock_cohere", "embed"):
+            try:
+                return self._delegate.embed_query(text)
+            except Exception as exc:
+                self._metrics.increment(
+                    PROVIDER_FAILURE,
+                    provider="bedrock_cohere",
+                    operation="embed",
+                    reason=_failure_reason(exc),
+                )
+                raise
+
+
+class InstrumentedReranker:
+    def __init__(self, delegate: Any, metrics: Metrics) -> None:
+        self._delegate = delegate
+        self._metrics = metrics
+
+    def rerank(
+        self, query: str, documents: list[str], top_n: int | None = None
+    ) -> list[int]:
+        with dependency_span("bedrock_cohere", "rerank"):
+            try:
+                return self._delegate.rerank(query, documents, top_n=top_n)
+            except Exception as exc:
+                self._metrics.increment(
+                    PROVIDER_FAILURE,
+                    provider="bedrock_cohere",
+                    operation="rerank",
+                    reason=_failure_reason(exc),
+                )
+                raise
+
+
+class InstrumentedRetrievalService(RetrievalService):
+    def __init__(self, *args, metrics: Metrics, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self._observability_metrics = metrics
+
+    def retrieve_framed(self, *args, **kwargs):
+        with stage("retrieval"):
+            section_key = args[1] if len(args) > 1 else kwargs.get("section_key")
+            is_overview = args[3] if len(args) > 3 else kwargs.get("is_overview", False)
+            route = "section" if section_key else (
+                "overview" if is_overview else "similarity"
+            )
+            self._observability_metrics.increment(RETRIEVAL_ROUTE, route=route)
+            try:
+                result = super().retrieve_framed(*args, **kwargs)
+            except Exception as exc:
+                self._record_retrieval_failure(exc)
+                raise
+            self._annotate_result(result)
+            return result
+
+    def retrieve(self, *args, **kwargs):
+        with stage("retrieval"):
+            self._observability_metrics.increment(RETRIEVAL_ROUTE, route="other")
+            try:
+                result = super().retrieve(*args, **kwargs)
+            except Exception as exc:
+                self._record_retrieval_failure(exc)
+                raise
+            self._annotate_result(result)
+            return result
+
+    def retrieve_by_indication(self, *args, **kwargs):
+        with stage("retrieval"):
+            self._observability_metrics.increment(RETRIEVAL_ROUTE, route="indication")
+            try:
+                result = super().retrieve_by_indication(*args, **kwargs)
+            except Exception as exc:
+                self._record_retrieval_failure(exc)
+                raise
+            self._annotate_result(result)
+            return result
+
+    def _record_retrieval_failure(self, exc: BaseException) -> None:
+        self._observability_metrics.increment(
+            PROVIDER_FAILURE,
+            provider="qdrant",
+            operation="retrieve",
+            reason=_failure_reason(exc),
+        )
+
+    @staticmethod
+    def _annotate_result(result) -> None:
+        annotate_current_span(
+            **{
+                "duocthu.decision": result.decision.value,
+                "duocthu.reason": result.reason,
+                "duocthu.evidence_count": len(result.evidence),
+            }
+        )
+
+    def _rerank(self, query, hits):
+        with stage("rerank", configured=self._reranker is not None):
+            return super()._rerank(query, hits)
+
+    def _decide(self, evidence, is_drug_overview=False):
+        with stage("evidence", evidence_count=len(evidence)):
+            return super()._decide(evidence, is_drug_overview=is_drug_overview)
+
+
+class InstrumentedGroundedAnswerService(GroundedAnswerService):
+    def _generate(self, *args, **kwargs):
+        with stage("generation"):
+            return super()._generate(*args, **kwargs)
+
+    def _verify_entailment(self, *args, **kwargs):
+        with stage("entailment"):
+            return super()._verify_entailment(*args, **kwargs)
+
+
+class InstrumentedQueryUnderstander:
+    def __init__(self, delegate: Any) -> None:
+        self._delegate = delegate
+
+    def understand(self, *args, **kwargs):
+        with stage("understanding"):
+            frame = self._delegate.understand(*args, **kwargs)
+            annotate_current_span(
+                **{
+                    "duocthu.turn_type": frame.turn_type,
+                    "duocthu.needs_clarify": frame.needs_clarify,
+                    "duocthu.system_error": frame.system_error,
+                }
+            )
+            return frame
+
+
+class InstrumentedRagAgent(RagAgent):
+    def __init__(self, *args, metrics: Metrics, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self._observability_metrics = metrics
+
+    def handle(self, *args, **kwargs):
+        reply = super().handle(*args, **kwargs)
+        if reply.decision == "clarify":
+            self._observability_metrics.increment(CLARIFY_ASKED, reason=reply.reason)
+        return reply
+
+    def _get_history(self, *args, **kwargs):
+        with stage("context"):
+            return super()._get_history(*args, **kwargs)
+
+    def _route(self, *args, **kwargs):
+        with stage("routing"):
+            return super()._route(*args, **kwargs)
+
+    def _remember(self, *args, **kwargs):
+        with stage("persistence"):
+            return super()._remember(*args, **kwargs)
@@ -18,18 +18,28 @@ from typing import Protocol
 class Metrics(Protocol):
    def increment(self, name: str, **labels: str) -> None: ...

+    def observe(self, name: str, value: float, **labels: str) -> None: ...
+

 class NullMetrics:
    def increment(self, name: str, **labels: str) -> None:  # noqa: ARG002
        # Deliberately inert: the default when no metrics stack is configured.
        return None

+    def observe(
+        self, name: str, value: float, **labels: str  # noqa: ARG002
+    ) -> None:
+        return None
+

 class InMemoryMetrics:
    """Reference implementation of the contract; also what tests assert on."""

    def __init__(self) -> None:
        self.counts: dict[tuple[str, tuple[tuple[str, str], ...]], int] = {}
+        self.observations: dict[
+            tuple[str, tuple[tuple[str, str], ...]], list[float]
+        ] = {}

    def increment(self, name: str, **labels: str) -> None:
        key = (name, tuple(sorted(labels.items())))
@@ -40,6 +50,22 @@ class InMemoryMetrics:
            return self.counts.get((name, tuple(sorted(labels.items()))), 0)
        return sum(count for (n, _), count in self.counts.items() if n == name)

+    def observe(self, name: str, value: float, **labels: str) -> None:
+        key = (name, tuple(sorted(labels.items())))
+        self.observations.setdefault(key, []).append(value)
+
+    def observed(self, name: str, **labels: str) -> tuple[float, ...]:
+        if labels:
+            return tuple(
+                self.observations.get((name, tuple(sorted(labels.items()))), [])
+            )
+        return tuple(
+            value
+            for (metric_name, _), values in self.observations.items()
+            if metric_name == name
+            for value in values
+        )
+

 RETRIEVAL_ROUTE = "duocthu_retrieval_route_total"
 ABSTENTION = "duocthu_abstention_total"
@@ -60,3 +86,12 @@ FOLLOWUP_INHERITED = "duocthu_followup_inherited_total"
 # degradation actually happens, since a silent fail-open with no counter is
 # indistinguishable from tracing quietly working.
 TRACE_WRITE_FAILED = "duocthu_trace_write_failed_total"
+
+# HTTP and request-pipeline telemetry. Every label is a bounded vocabulary:
+# route is a route template (never a raw path), status is a class (2xx/4xx/5xx),
+# and stage/provider/reason values are normalized by the Prometheus adapter.
+REQUESTS = "duocthu_requests_total"
+REQUEST_DURATION = "duocthu_request_duration_seconds"
+STAGE_DURATION = "duocthu_stage_duration_seconds"
+DECISION = "duocthu_decision_total"
+PROVIDER_FAILURE = "duocthu_provider_failure_total"
@@ -0,0 +1,217 @@
+"""Request correlation, OpenTelemetry spans and stage timing.
+
+This module deliberately has a no-op default. The RAG domain remains runnable
+without the OpenTelemetry packages or collector, while a configured runtime
+gets one trace context shared by FastAPI's async middleware and its sync
+thread-pool endpoint.
+"""
+from __future__ import annotations
+
+import re
+import uuid
+from contextlib import contextmanager
+from contextvars import ContextVar
+from functools import wraps
+from time import perf_counter
+from typing import Any, Iterator, Mapping
+
+from .metrics import Metrics, NullMetrics, STAGE_DURATION
+
+_CORRELATION_ID: ContextVar[str | None] = ContextVar(
+    "duocthu_correlation_id", default=None
+)
+_STAGE: ContextVar[str | None] = ContextVar("duocthu_stage", default=None)
+_SAFE_CORRELATION = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{0,127}$")
+_metrics: Metrics = NullMetrics()
+
+try:
+    from opentelemetry import trace
+
+    _tracer = trace.get_tracer("duocthu.ai-service")
+except ImportError:  # pragma: no cover - exercised only in minimal installs
+    trace = None
+    _tracer = None
+
+
+def configure_telemetry(settings: Any, metrics: Metrics | None = None) -> bool:
+    """Configure one OTLP tracer provider; return whether tracing is active.
+
+    Metrics are configured independently and continue to work when tracing is
+    disabled or optional OpenTelemetry packages are absent.
+    """
+    global _metrics, _tracer
+    _metrics = metrics or NullMetrics()
+    if not getattr(settings, "otel_enabled", False):
+        return False
+
+    try:
+        from opentelemetry import trace as otel_trace
+        from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+        from opentelemetry.sdk.resources import Resource
+        from opentelemetry.sdk.trace import TracerProvider
+        from opentelemetry.sdk.trace.export import BatchSpanProcessor
+        from opentelemetry.sdk.trace.sampling import ParentBased, TraceIdRatioBased
+    except ImportError:
+        return False
+
+    current = otel_trace.get_tracer_provider()
+    # OpenTelemetry only permits setting the global provider once. Reuse a
+    # provider installed by a host/test instead of replacing it and warning.
+    if isinstance(current, TracerProvider):
+        _tracer = current.get_tracer("duocthu.ai-service")
+        return True
+
+    provider = TracerProvider(
+        resource=Resource.create(
+            {
+                "service.name": getattr(settings, "otel_service_name", "ai-service"),
+                "service.version": "0.1.0",
+                "deployment.environment": getattr(settings, "environment", "local"),
+            }
+        ),
+        sampler=ParentBased(
+            TraceIdRatioBased(float(getattr(settings, "otel_sample_ratio", 1.0)))
+        ),
+    )
+    exporter = OTLPSpanExporter(endpoint=settings.otel_exporter_otlp_endpoint)
+    provider.add_span_processor(BatchSpanProcessor(exporter))
+    otel_trace.set_tracer_provider(provider)
+    _tracer = provider.get_tracer("duocthu.ai-service")
+    return True
+
+
+def normalize_correlation_id(candidate: str | None) -> str:
+    value = (candidate or "").strip()
+    return value if _SAFE_CORRELATION.fullmatch(value) else str(uuid.uuid4())
+
+
+@contextmanager
+def correlation_context(candidate: str | None) -> Iterator[str]:
+    correlation_id = normalize_correlation_id(candidate)
+    token = _CORRELATION_ID.set(correlation_id)
+    try:
+        yield correlation_id
+    finally:
+        _CORRELATION_ID.reset(token)
+
+
+def current_correlation_id() -> str:
+    value = _CORRELATION_ID.get()
+    return value or str(uuid.uuid4())
+
+
+def current_trace_id() -> str | None:
+    if trace is None:
+        return None
+    context = trace.get_current_span().get_span_context()
+    if not context.is_valid:
+        return None
+    return f"{context.trace_id:032x}"
+
+
+def current_stage() -> str | None:
+    return _STAGE.get()
+
+
+@contextmanager
+def request_span(
+    method: str, route: str, headers: Mapping[str, str]
+) -> Iterator[Any]:
+    """Start a server span and extract an incoming W3C ``traceparent``."""
+    if _tracer is None:
+        yield None
+        return
+    try:
+        from opentelemetry import propagate
+        from opentelemetry.trace import SpanKind
+    except ImportError:  # pragma: no cover - minimal install fallback
+        yield None
+        return
+
+    parent = propagate.extract(headers)
+    with _tracer.start_as_current_span(
+        f"{method} {route}",
+        context=parent,
+        kind=SpanKind.SERVER,
+        attributes={
+            "http.request.method": method,
+            "http.route": route,
+            "duocthu.correlation_id": current_correlation_id(),
+        },
+    ) as span:
+        yield span
+
+
+@contextmanager
+def dependency_span(provider: str, operation: str) -> Iterator[Any]:
+    """Trace one external dependency call without creating another histogram."""
+    if _tracer is None:
+        yield None
+        return
+    with _tracer.start_as_current_span(
+        f"provider.{provider}.{operation}",
+        attributes={
+            "duocthu.provider": provider,
+            "duocthu.operation": operation,
+        },
+    ) as span:
+        yield span
+
+
+@contextmanager
+def stage(name: str, **attributes: Any) -> Iterator[Any]:
+    """Create a child span and observe a low-cardinality stage histogram."""
+    started = perf_counter()
+    outcome = "ok"
+    span = None
+    token = _STAGE.set(name)
+    try:
+        if _tracer is None:
+            yield None
+        else:
+            safe_attributes = {
+                key: value
+                for key, value in attributes.items()
+                if value is not None and isinstance(value, (str, bool, int, float))
+            }
+            safe_attributes["duocthu.stage"] = name
+            with _tracer.start_as_current_span(
+                f"rag.stage.{name}", attributes=safe_attributes
+            ) as span:
+                yield span
+    except BaseException as exc:
+        outcome = "cancelled" if exc.__class__.__name__ == "CancelledError" else "error"
+        if span is not None:
+            span.set_attribute("duocthu.outcome", outcome)
+        raise
+    finally:
+        elapsed = perf_counter() - started
+        if span is not None:
+            span.set_attribute("duocthu.outcome", outcome)
+            span.set_attribute("duocthu.duration_ms", elapsed * 1000)
+        _metrics.observe(STAGE_DURATION, elapsed, stage=name, outcome=outcome)
+        _STAGE.reset(token)
+
+
+def annotate_current_span(**attributes: Any) -> None:
+    if trace is None:
+        return
+    span = trace.get_current_span()
+    if not span.is_recording():
+        return
+    for key, value in attributes.items():
+        if value is not None and isinstance(value, (str, bool, int, float)):
+            span.set_attribute(key, value)
+
+
+def traced_stage(name: str):
+    """Decorator form for pure functions that should remain otherwise untouched."""
+    def decorate(function):
+        @wraps(function)
+        def wrapped(*args, **kwargs):
+            with stage(name):
+                return function(*args, **kwargs)
+
+        return wrapped
+
+    return decorate
@@ -7,9 +7,15 @@ from fastapi import APIRouter, Depends, HTTPException, Request
 from pydantic import BaseModel, Field

 from rag.answer import GroundedAnswerService
-from rag.metrics import TRACE_WRITE_FAILED, Metrics, NullMetrics
+from rag.metrics import DECISION, TRACE_WRITE_FAILED, Metrics, NullMetrics
 from rag.models import QueryIntent, SubjectScope
 from rag.policy import resolve_subject_scope
+from rag.telemetry import (
+    annotate_current_span,
+    current_correlation_id,
+    current_trace_id,
+    stage,
+)


 class TraceWriter(Protocol):
@@ -61,6 +67,8 @@ class AnswerPlanResponse(BaseModel):

 class RagQueryResponse(BaseModel):
    trace_id: str
+    correlation_id: str
+    otel_trace_id: str | None = None
    decision: str
    reason: str
    answer: str | None
@@ -166,7 +174,10 @@ def query_rag(
    traces: Annotated[TraceWriter, Depends(_trace_writer)],
    metrics: Annotated[Metrics, Depends(_metrics)],
 ) -> RagQueryResponse:
-    agent = getattr(request.app.state, "conversational", None)
+    with stage("receive"):
+        agent = getattr(request.app.state, "conversational", None)
+        subject_scope = resolve_subject_scope(payload.query, payload.subject_scope)
+        intent = payload.intent

    # `payload.subject_scope`/`payload.intent` are what the CALLER claims —
    # logged below for audit, but the RagAgent path does not take them as an
@@ -175,9 +186,6 @@ def query_rag(
    # own LLM understanding call, and does not gate on intent at all (this
    # product is for doctors/pharmacists; a client label must not be, and
    # here structurally cannot be, the safety decision — F-02).
-    subject_scope = resolve_subject_scope(payload.query, payload.subject_scope)
-    intent = payload.intent
-
    if agent is not None:
        # The live path (F-03): one LLM call understands the turn (drug
        # identity against the real catalog, turn type, population/weight),
@@ -199,7 +207,8 @@ def query_rag(
        # No generator configured (ANSWER_PROVIDER=disabled): there is no LLM
        # to understand a turn with, so this is retrieval-only, single-turn,
        # unchanged from before F-03.
-        grounded = answers.answer(payload.query, subject_scope, intent)
+        with stage("routing"):
+            grounded = answers.answer(payload.query, subject_scope, intent)
        if grounded.clarification is not None:
            decision, reason = "clarify", "needs_more_info"
            answer = grounded.clarification
@@ -229,32 +238,51 @@ def query_rag(
    # with whether the answer was safe. `trace_id` degrades to a local,
    # unpersisted uuid — still a valid response field, just not one `GET
    # /v1/rag/trace/{id}` (if it existed) could later look up.
+    correlation_id = current_correlation_id()
+    otel_trace_id = current_trace_id()
    try:
-        trace_id = traces.save(
-            query=payload.query,
-            # The resolved (server-derived) values, not the caller's claim —
-            # this is what actually gated the answer, so it's what the trace
-            # must show.
-            subject_scope=subject_scope.value,
-            intent=intent.value,
-            decision=decision,
-            reason=reason,
-            resolved_drug_id=resolved_drug_id,
-            citations=tuple(item.model_dump() for item in citations),
-        )
+        with stage("persistence"):
+            trace_id = traces.save(
+                query=payload.query,
+                # The resolved (server-derived) values, not the caller's claim —
+                # this is what actually gated the answer, so it's what the trace
+                # must show.
+                subject_scope=subject_scope.value,
+                intent=intent.value,
+                decision=decision,
+                reason=reason,
+                resolved_drug_id=resolved_drug_id,
+                citations=tuple(item.model_dump() for item in citations),
+                correlation_id=correlation_id,
+                otel_trace_id=otel_trace_id,
+            )
    except Exception:
        metrics.increment(TRACE_WRITE_FAILED)
        trace_id = str(uuid.uuid4())
-    return RagQueryResponse(
-        trace_id=trace_id,
-        decision=decision,
-        reason=reason,
-        answer=answer,
-        resolved_drug_id=resolved_drug_id,
-        citations=citations,
-        generated=generated,
-        quick_replies=quick_replies,
-        blocks=blocks,
-        answer_mode=answer_mode,
-        answer_plan=answer_plan,
+    metrics.increment(DECISION, decision=decision, reason=reason)
+    annotate_current_span(
+        **{
+            "duocthu.decision": decision,
+            "duocthu.reason": reason,
+            "duocthu.citation_count": len(citations),
+            "duocthu.generated": generated,
+            "duocthu.persisted_trace_id": trace_id,
+        }
    )
+    with stage("response"):
+        response = RagQueryResponse(
+            trace_id=trace_id,
+            correlation_id=correlation_id,
+            otel_trace_id=otel_trace_id,
+            decision=decision,
+            reason=reason,
+            answer=answer,
+            resolved_drug_id=resolved_drug_id,
+            citations=citations,
+            generated=generated,
+            quick_replies=quick_replies,
+            blocks=blocks,
+            answer_mode=answer_mode,
+            answer_plan=answer_plan,
+        )
+    return response
@@ -1,5 +1,6 @@
 from fastapi.testclient import TestClient

+from adapters.prometheus import PrometheusMetrics
 from config import Settings
 from main import create_app
 from rag.agent import AgentReply
@@ -177,3 +178,63 @@ def test_a_trace_write_failure_does_not_turn_a_good_answer_into_a_500():
    assert body["answer"] == "Liều 500 mg [1]."
    assert body["trace_id"]  # a locally-generated fallback id, still present
    assert metrics.total(TRACE_WRITE_FAILED) == 1
+
+
+class _AnyRouting:
+    def retrieve(self, query, subject_scope, intent):
+        return RetrievalResult(EvidenceDecision.ABSTAIN, "drug_not_resolved")
+
+
+def test_correlation_id_round_trips_through_headers_body_and_trace_row():
+    traces = MemoryTraceWriter()
+    app = create_app(
+        settings=Settings(embedding_provider="disabled"),
+        answer_service=GroundedAnswerService(_AnyRouting()),
+        trace_writer=traces,
+    )
+
+    response = TestClient(app).post(
+        "/v1/rag/query",
+        headers={"X-Correlation-ID": "req-test-1"},
+        json={
+            "query": "Paracetamol dose?",
+            "subject_scope": "human",
+            "intent": "fact_lookup",
+        },
+    )
+
+    assert response.status_code == 200
+    assert response.headers["x-correlation-id"] == "req-test-1"
+    assert response.json()["correlation_id"] == "req-test-1"
+    assert traces.rows[0]["correlation_id"] == "req-test-1"
+    assert "otel_trace_id" in traces.rows[0]
+
+
+def test_metrics_endpoint_exposes_request_decision_and_stage_histograms():
+    metrics = PrometheusMetrics()
+    app = create_app(
+        settings=Settings(embedding_provider="disabled"),
+        answer_service=GroundedAnswerService(_AnyRouting()),
+        trace_writer=MemoryTraceWriter(),
+        metrics=metrics,
+    )
+    client = TestClient(app)
+
+    response = client.post(
+        "/v1/rag/query",
+        json={
+            "query": "Paracetamol dose?",
+            "subject_scope": "human",
+            "intent": "fact_lookup",
+        },
+    )
+    assert response.status_code == 200
+
+    scrape = client.get("/metrics")
+    assert scrape.status_code == 200
+    body = scrape.text
+    assert 'duocthu_requests_total{method="POST",route="/v1/rag/query",status="2xx"}' in body
+    assert 'duocthu_decision_total{decision="abstain",reason="drug_not_resolved"}' in body
+    assert 'duocthu_stage_duration_seconds_count{outcome="ok",stage="receive"}' in body
+    assert 'duocthu_stage_duration_seconds_count{outcome="ok",stage="persistence"}' in body
+    assert 'duocthu_stage_duration_seconds_count{outcome="ok",stage="response"}' in body
@@ -11,7 +11,7 @@ import json
 import pytest

 from rag import grounding
-from rag.answer import GroundedAnswerService
+from rag.answer import DISCLAIMER, GroundedAnswerService
 from rag.budget import RequestBudgetExhausted
 from rag.metrics import GENERATION_REJECTED, GENERATION_SERVED, InMemoryMetrics
 from rag.models import (
@@ -116,6 +116,62 @@ def _answer(payload, result: RetrievalResult | None = None, entailment_payload=N
    return grounded, metrics


+# --- the disclaimer guardrail -------------------------------------------------
+# `docs/architecture.md` specifies the disclaimer at several layers. Only the
+# web banner existed; `packages/shared-types/src/dto/chat.ts` declared the
+# field but nothing filled it, so a consumer other than that one UI got medical
+# content with nothing attached. It is a dataclass default rather than
+# something each call site adds, so these tests are about the paths that could
+# plausibly skip it: rejections, abstains and clarifications.
+
+
+def test_a_served_answer_carries_the_disclaimer():
+    grounded, _ = _answer(
+        {"claims": [{"text": "Người lớn uống 500 mg", "citations": [1]}],
+         "evidence_sufficient": True},
+    )
+
+    assert grounded.generated is True
+    assert grounded.disclaimer == DISCLAIMER
+    assert grounded.disclaimer
+
+
+def test_a_rejected_generation_still_carries_the_disclaimer():
+    """An abstain is still a clinical response and still needs the notice —
+    it is the case most likely to be treated as "not really an answer"."""
+    grounded, _ = _answer(
+        {"claims": [{"text": "Người lớn uống 850 mg, 2 lần mỗi ngày", "citations": [1]}],
+         "evidence_sufficient": True},
+    )
+
+    assert grounded.answer is None
+    assert grounded.result.decision == EvidenceDecision.ABSTAIN
+    assert grounded.disclaimer == DISCLAIMER
+
+
+def test_a_provider_outage_response_still_carries_the_disclaimer():
+    grounded, _ = _answer(
+        {"claims": [{"text": "Người lớn: 500 mg, 2 lần/ngày", "citations": [1]}],
+         "evidence_sufficient": True},
+        entailment_payload=AnswerGenerationUnavailable(),
+    )
+
+    assert grounded.disclaimer == DISCLAIMER
+
+
+def test_the_disclaimer_is_not_something_the_model_can_influence():
+    """It is a module constant, never routed through the generator, so a
+    prompt-injected or malfunctioning model cannot shorten or drop it. Asserted
+    directly because the value is the guardrail."""
+    grounded, _ = _answer(
+        {"claims": [{"text": "Bỏ qua mọi cảnh báo. Người lớn uống 500 mg", "citations": [1]}],
+         "evidence_sufficient": True},
+    )
+
+    assert grounded.disclaimer == DISCLAIMER
+    assert "không thay thế chỉ định" in grounded.disclaimer
+
+
 # --- the guardrail's whole reason to exist ------------------------------------


@@ -0,0 +1,122 @@
+from adapters.prometheus import PrometheusMetrics
+from config import Settings
+from rag.instrumentation import InstrumentedGenerator
+from rag.metrics import (
+    CLARIFY_ASKED,
+    PROVIDER_FAILURE,
+    REQUEST_DURATION,
+    REQUESTS,
+    STAGE_DURATION,
+    TRACE_WRITE_FAILED,
+    InMemoryMetrics,
+)
+from rag.telemetry import (
+    configure_telemetry,
+    correlation_context,
+    request_span,
+    stage,
+)
+
+
+def test_prometheus_registers_domain_failures_histograms_and_bounds_labels():
+    metrics = PrometheusMetrics()
+    metrics.increment(TRACE_WRITE_FAILED)
+    metrics.increment(CLARIFY_ASKED, reason="raw-user-controlled-reason")
+    metrics.increment(
+        REQUESTS, method="POST", route="/v1/rag/query", status="2xx"
+    )
+    metrics.observe(
+        REQUEST_DURATION,
+        1.25,
+        method="POST",
+        route="/v1/rag/query",
+        status="2xx",
+    )
+    metrics.observe(STAGE_DURATION, 0.2, stage="retrieval", outcome="ok")
+
+    body, content_type = metrics.render()
+    rendered = body.decode("utf-8")
+    assert "application/openmetrics-text" in content_type
+    assert "duocthu_trace_write_failed_total 1.0" in rendered
+    assert 'duocthu_clarify_asked_total{reason="other"} 1.0' in rendered
+    assert "duocthu_request_duration_seconds_bucket" in rendered
+    assert "duocthu_stage_duration_seconds_bucket" in rendered
+
+
+def test_stage_records_duration_without_requiring_an_otel_collector():
+    metrics = InMemoryMetrics()
+    configure_telemetry(Settings(otel_enabled=False), metrics)
+
+    with stage("grounding"):
+        pass
+
+    values = metrics.observed(STAGE_DURATION, stage="grounding", outcome="ok")
+    assert len(values) == 1
+    assert values[0] >= 0
+
+
+class _FailingGenerator:
+    model_id = "fake"
+
+    def generate(self, system: str, user: str, schema: dict) -> str:
+        raise ConnectionError("provider offline")
+
+
+def test_provider_failure_uses_bounded_classification_not_exception_text():
+    metrics = InMemoryMetrics()
+    generator = InstrumentedGenerator(_FailingGenerator(), metrics)
+
+    try:
+        generator.generate("system", "user", {})
+    except ConnectionError:
+        pass
+    else:  # pragma: no cover - assertion guard
+        raise AssertionError("expected provider failure")
+
+    assert metrics.total(
+        PROVIDER_FAILURE,
+        provider="other",
+        operation="generate",
+        reason="provider_unavailable",
+    ) == 1
+
+
+def test_otel_server_and_stage_spans_share_trace_and_emit_metric_exemplar():
+    otel_trace = pytest.importorskip("opentelemetry.trace")
+    sdk_trace = pytest.importorskip("opentelemetry.sdk.trace")
+    exporter_module = pytest.importorskip(
+        "opentelemetry.sdk.trace.export.in_memory_span_exporter"
+    )
+    export_module = pytest.importorskip("opentelemetry.sdk.trace.export")
+
+    exporter = exporter_module.InMemorySpanExporter()
+    provider = sdk_trace.TracerProvider()
+    provider.add_span_processor(export_module.SimpleSpanProcessor(exporter))
+    otel_trace.set_tracer_provider(provider)
+
+    metrics = PrometheusMetrics()
+    configure_telemetry(
+        Settings(
+            otel_enabled=True,
+            otel_exporter_otlp_endpoint="http://127.0.0.1:4318/v1/traces",
+        ),
+        metrics,
+    )
+    with correlation_context("req-span-test"):
+        with request_span("POST", "/v1/rag/query", {}):
+            with stage("retrieval"):
+                pass
+
+    provider.force_flush()
+    spans = exporter.get_finished_spans()
+    assert {item.name for item in spans} == {
+        "POST /v1/rag/query",
+        "rag.stage.retrieval",
+    }
+    assert len({item.context.trace_id for item in spans}) == 1
+    assert spans[-1].attributes["duocthu.correlation_id"] == "req-span-test"
+
+    rendered = metrics.render()[0].decode("utf-8")
+    assert "duocthu_stage_duration_seconds_bucket" in rendered
+    assert 'trace_id="' in rendered
+import pytest