apecloud
diff --git a/‎.github/workflows/e2e-http-smoke.yml‎
Lines changed: 2 additions & 2 deletions b/‎.github/workflows/e2e-http-smoke.yml‎
Lines changed: 2 additions & 2 deletions
diff --git a/‎Makefile‎
Lines changed: 6 additions & 13 deletions b/‎Makefile‎
Lines changed: 6 additions & 13 deletions
diff --git a/‎aperag/app.py‎
Lines changed: 122 additions & 3 deletions b/‎aperag/app.py‎
Lines changed: 122 additions & 3 deletions
@@ -65,7 +65,7 @@ jobs:
         if: failure()
         run: |
           docker compose -f docker-compose.yml ps || true
-          docker compose -f docker-compose.yml logs --no-color api celeryworker celerybeat postgres redis qdrant es || true
+          docker compose -f docker-compose.yml logs --no-color api postgres redis qdrant es || true
 
       - name: Stop Compose stack
         if: always()
@@ -170,7 +170,7 @@ jobs:
         if: failure()
         run: |
           docker compose -f docker-compose.yml ps || true
-          docker compose -f docker-compose.yml logs --no-color api celeryworker celerybeat postgres redis qdrant es || true
+          docker compose -f docker-compose.yml logs --no-color api postgres redis qdrant es || true
 
       - name: Stop Compose stack
         if: always()
 
@@ -39,9 +39,6 @@ help:
 	@printf "  make stack-logs           Tail stack logs\n\n"
 	@printf "Services\n"
 	@printf "  make serve-api            Run backend API locally\n"
-	@printf "  make serve-worker         Run celery worker locally\n"
-	@printf "  make serve-beat           Run celery beat locally\n"
-	@printf "  make serve-flower         Run flower locally\n"
 	@printf "  make serve-web            Run frontend locally\n\n"
 	@printf "Tests\n"
 	@printf "  make test-all             Run unit + integration + pytest E2E suites\n"
@@ -171,19 +168,15 @@ stack-logs:
 ##################################################
 
 # Local development services
-.PHONY: serve-api serve-web serve-worker serve-flower serve-beat
+# Wave 3 T3.1 chunk 3: ``serve-worker`` / ``serve-beat`` / ``serve-flower``
+# targets removed alongside the Celery infrastructure deletion. The
+# in-process ``aperag.indexing`` runtime (worker pool + reconciler +
+# cleanup loops) is spawned by the FastAPI lifespan when ``serve-api``
+# starts, so no separate worker / beat / monitoring command is needed.
+.PHONY: serve-api serve-web
 serve-api: db-migrate
 	uvicorn aperag.app:app --host 0.0.0.0 --log-config scripts/uvicorn-log-config.yaml
 
-serve-worker:
-	celery -A config.celery worker -B -l INFO --pool=threads --concurrency=16
-
-serve-beat:
-	celery -A config.celery beat -l INFO
-
-serve-flower:
-	celery -A config.celery flower --conf/flowerconfig.py
-
 serve-web:
 	cd ./web && yarn dev
 
 
@@ -13,6 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import asyncio  # noqa: E402
+
 from aperag.config import settings
 from aperag.observability import (
     bind_observability_context,
@@ -205,9 +207,126 @@ async def initialize_user_quota(self, user_id: str) -> None:
 
 
 async def combined_lifespan(app: FastAPI):
-    """Combined lifespan manager for the API and MCP server."""
-    async with mcp_app.lifespan(app):
-        yield
+    """Combined lifespan manager for the API + MCP server + indexing runtime.
+
+    The indexing runtime (Wave 3 T3.1 wire-in) launches the per-modality
+    worker pool + reconciler + cleanup loop only when
+    ``settings.indexing_mode == "async"``. In ``inline`` mode the
+    upload-side ``dispatch_indexing(mode=INLINE)`` runs derive + sync +
+    cutover within the request coroutine, so no background workers are
+    needed (per design pack §L Tier-1 deployment).
+
+    The runtime is started as background asyncio tasks (not subprocesses)
+    so a single FastAPI process owns its workers — matches the §E.2
+    "one Python process per modality" architecture for the in-process
+    deployment topology. Tier-3 horizontal scale-out runs separate
+    worker processes; that wiring lives in a future ops launcher.
+    """
+    indexing_runtime_tasks: list[asyncio.Task[None]] = []
+    indexing_shutdown: asyncio.Event | None = None
+
+    if settings.indexing_mode == "async":
+        # Lazy imports — pulling the indexing runtime symbols at app
+        # start-up time keeps ``aperag/app.py`` cold-start fast and
+        # confines the import surface to the wired branch.
+        from aperag.config import sync_engine
+        from aperag.indexing import (
+            InMemoryWorkQueue,
+            run_cleanup_loop,
+            run_fulltext_worker,
+            run_graph_worker,
+            run_reconcile_loop,
+            run_summary_worker,
+            run_vector_worker,
+            run_vision_worker,
+        )
+
+        indexing_shutdown = asyncio.Event()
+        # Single process-local InMemoryWorkQueue is the default
+        # transport for the in-process topology. Tier-3 production
+        # swaps this for a Redis-backed WorkQueue (RPUSH / BLPOP) by
+        # injecting via app state at deploy time — Wave 3 follow-up.
+        queue = InMemoryWorkQueue()
+        engine = sync_engine
+
+        # Worker factory — per-task lazy construction. The async
+        # worker entrypoints (``run_*_worker``) call this closure on
+        # every BLPOP'd payload to materialise the concrete
+        # :class:`ModalityWorker` for that ``(collection, modality)``
+        # pair. ``ProductionWorkerFactory`` resolves the collection
+        # row, picks the right backend (Qdrant / Elasticsearch +
+        # configured embedder / completion model), and constructs the
+        # worker — all backed by the existing build helpers
+        # (``get_collection_embedding_service_sync`` /
+        # ``get_vector_db_connector`` / ``get_object_store``) so this
+        # is composition, not re-implementation. Construction failures
+        # raise :class:`WorkerFactoryError`; the orchestrator runner
+        # catches that and finalises the row FAILED so §I.2 retry
+        # picks it up. Per architect msg=7782ebe0.
+        from aperag.indexing.worker_factory import ProductionWorkerFactory
+
+        worker_factory = ProductionWorkerFactory(engine=engine)
+
+        worker_kwargs = dict(
+            engine=engine,
+            queue=queue,
+            worker_factory=worker_factory,
+            shutdown=indexing_shutdown,
+        )
+        indexing_runtime_tasks.append(asyncio.create_task(run_vector_worker(**worker_kwargs)))
+        indexing_runtime_tasks.append(asyncio.create_task(run_fulltext_worker(**worker_kwargs)))
+        indexing_runtime_tasks.append(asyncio.create_task(run_graph_worker(**worker_kwargs)))
+        indexing_runtime_tasks.append(asyncio.create_task(run_summary_worker(**worker_kwargs)))
+        indexing_runtime_tasks.append(asyncio.create_task(run_vision_worker(**worker_kwargs)))
+        indexing_runtime_tasks.append(
+            asyncio.create_task(
+                run_reconcile_loop(
+                    engine=engine,
+                    queue=queue,
+                    shutdown=indexing_shutdown,
+                )
+            )
+        )
+        indexing_runtime_tasks.append(
+            asyncio.create_task(
+                run_cleanup_loop(
+                    engine=engine,
+                    workers={},  # T3.3 follow-up: pass concrete worker registry
+                    shutdown=indexing_shutdown,
+                )
+            )
+        )
+
+        # Stash on app state so request handlers can dispatch via the
+        # same queue / engine the workers consume.
+        app.state.indexing_queue = queue
+        app.state.indexing_engine = engine
+
+        # Service-layer callers (aperag/domains/**) consume the same
+        # triple through the process-wide IndexingRuntime singleton —
+        # they don't have a Request handle for app.state. Workers map
+        # is empty in the async-default deployment; T3.3 follow-up
+        # populates concrete factories per modality.
+        from aperag.indexing.runtime import IndexingRuntime, set_runtime
+
+        set_runtime(IndexingRuntime(engine=engine, queue=queue, workers={}))
+    else:
+        app.state.indexing_queue = None
+        app.state.indexing_engine = None
+        from aperag.indexing.runtime import set_runtime
+
+        set_runtime(None)
+
+    try:
+        async with mcp_app.lifespan(app):
+            yield
+    finally:
+        if indexing_shutdown is not None:
+            indexing_shutdown.set()
+        if indexing_runtime_tasks:
+            # Drain in-flight worker / reconciler / cleanup loops with
+            # a short grace window so a SIGTERM does not abort mid-task.
+            await asyncio.gather(*indexing_runtime_tasks, return_exceptions=True)
 
 
 # Create the main FastAPI app with combined lifespan