CI fixes

.github/workflows/ci-secret.yaml

@@ -63,6 +63,18 @@ jobs:
       run: |
         make docker-up
 
+    - name: Wait for graph readiness
+      run: |
+        echo "Waiting for graph to finish initializing..."
+        for i in $(seq 1 360); do
+          if curl -sf http://localhost:8000/conversations/ready | grep -q '"ready"'; then
+            echo "Graph is ready"
+            break
+          fi
+          echo "Waiting for graph initialization... ($i/180)"
+          sleep 10
+        done
+
     - name: Run LLM CI
       id: llm_tests
       working-directory: evaluation

backend/src/api/main.py

@@ -11,9 +11,12 @@
 
 @asynccontextmanager
 async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
-    """Initialize database on startup."""
+    """Initialize database on startup and start graph background init."""
     logger.info("Initializing database connection...")
     init_database()
+    logger.info("Starting graph initialization in background...")
+    from .routers.conversations import start_graph_init
+    start_graph_init()
     yield
     logger.info("Shutting down...")
 

backend/src/api/routers/conversations.py

@@ -1,8 +1,9 @@
 import os
 import logging
+import threading
 from dotenv import load_dotenv
 
-from typing import Any
+from typing import Any, Optional
 from uuid import UUID
 from fastapi import APIRouter, Depends, HTTPException
 from langchain_google_vertexai import ChatVertexAI
@@ -204,17 +205,56 @@ def parse_agent_output(output: list) -> tuple[str, list[ContextSource], list[str
     return llm_response, context_sources, tools
 
 
-rg = RetrieverGraph(
-    llm_model=llm,
-    embeddings_config=embeddings_config,
-    reranking_model_name=hf_reranker,
-    use_cuda=use_cuda,
-    inbuilt_tool_calling=True,
-    fast_mode=fast_mode,
-    debug=debug,
-    enable_mcp=enable_mcp,
-)
-rg.initialize()
+_rg: Optional[RetrieverGraph] = None
+_rg_started = threading.Event()
+_rg_ready = threading.Event()
+
+
+def get_graph() -> Optional[RetrieverGraph]:
+    """Return the initialized graph, or None if not ready yet."""
+    return _rg if _rg_ready.is_set() else None
+
+
+def _initialize_graph() -> None:
+    """Build and initialize the RetrieverGraph (runs in background thread)."""
+    global _rg
+    graph = RetrieverGraph(
+        llm_model=llm,
+        embeddings_config=embeddings_config,
+        reranking_model_name=hf_reranker,
+        use_cuda=use_cuda,
+        inbuilt_tool_calling=True,
+        fast_mode=fast_mode,
+        debug=debug,
+        enable_mcp=enable_mcp,
+    )
+    graph.initialize()
+    _rg = graph
+    _rg_ready.set()
+
+
+def start_graph_init() -> None:
+    """Start graph initialization in a background thread (idempotent)."""
+    if _rg_started.is_set():
+        return
+    _rg_started.set()
+    threading.Thread(target=_initialize_graph, daemon=True).start()
+
+
+def reset_graph_state_for_testing() -> None:
+    """Reset graph state so tests can simulate a fresh startup."""
+    global _rg
+    _rg = None
+    _rg_started.clear()
+    _rg_ready.clear()
+
+
+@router.get("/ready")
+async def ready() -> dict[str, str]:
+    """Readiness probe — returns 'ready' when the graph is fully initialized."""
+    if _rg_ready.is_set():
+        return {"status": "ready"}
+    return {"status": "initializing"}
 
 
 chat_history: dict[UUID, list[dict[str, str]]] = {}
@@ -283,10 +323,11 @@ async def get_agent_response(
         "chat_history": get_history_str(db, conversation_uuid),
     }
 
-    if rg.graph is not None:
-        output = list(rg.graph.stream(inputs, stream_mode="updates"))
+    graph = get_graph()
+    if graph is not None and graph.graph is not None:
+        output = list(graph.graph.stream(inputs, stream_mode="updates"))
     else:
-        raise ValueError("RetrieverGraph not initialized.")
+        raise HTTPException(status_code=503, detail="Graph is still initializing. Please retry shortly.")
 
     llm_response, context_sources, tools = parse_agent_output(output)
 
@@ -382,8 +423,9 @@ async def get_response_stream(user_input: UserInput, db: Session | None) -> Any:
     current_llm_call_count = 1
     chunks: list[str] = []
 
-    if rg.graph is not None:
-        async for event in rg.graph.astream_events(inputs, version="v2"):
+    graph = get_graph()
+    if graph is not None and graph.graph is not None:
+        async for event in graph.graph.astream_events(inputs, version="v2"):
             chunk = event["event"]
 
             if chunk == "on_chat_model_end":
@@ -406,6 +448,9 @@ async def get_response_stream(user_input: UserInput, db: Session | None) -> Any:
                 if msg:
                     chunks.append(str(msg))
                 yield str(msg) + "\n\n"
+    else:
+        yield "Error: Graph is still initializing. Please retry shortly.\n\n"
+        return
 
     urls = list(set(urls))
     yield f"Sources: {', '.join(urls)}\n\n"

backend/src/vectorstores/faiss.py

@@ -1,4 +1,5 @@
 import os
+import time
 import logging
 from typing import Optional, Union
 from dotenv import load_dotenv
@@ -74,13 +75,13 @@ def faiss_db(self) -> Optional[FAISS]:
 
     @retry(
         stop=stop_after_attempt(5),
-        wait=wait_exponential(multiplier=2, min=10, max=120),
+        wait=wait_exponential(multiplier=2, min=60, max=600),
         retry=retry_if_exception(
             lambda e: "RESOURCE_EXHAUSTED" in str(e) or "429" in str(e)
         ),
         reraise=True,
     )
-    def _add_to_db(self, documents: list[Document]) -> None:
+    def _embed_and_add(self, documents: list[Document]) -> None:
         if self._faiss_db is None:
             self._faiss_db = FAISS.from_documents(
                 documents=documents,
@@ -90,6 +91,13 @@ def _add_to_db(self, documents: list[Document]) -> None:
         else:
             self._faiss_db.add_documents(documents)
 
+    def _add_to_db(self, documents: list[Document], batch_size: int = 100) -> None:
+        for i in range(0, len(documents), batch_size):
+            batch = documents[i : i + batch_size]
+            self._embed_and_add(batch)
+            if i + batch_size < len(documents):
+                time.sleep(1)
+
     def add_md_docs(
         self, folder_paths: list[str], chunk_size: int = 500, return_docs: bool = False
     ) -> Optional[list[Document]]:
@@ -229,7 +237,7 @@ def get_documents(self) -> list[Document]:
 
     @retry(
         stop=stop_after_attempt(5),
-        wait=wait_exponential(multiplier=2, min=10, max=120),
+        wait=wait_exponential(multiplier=2, min=60, max=600),
         retry=retry_if_exception(
             lambda e: "RESOURCE_EXHAUSTED" in str(e) or "429" in str(e)
         ),

backend/tests/test_api_conversations_streaming.py

@@ -37,7 +37,13 @@ def mock_retriever_graph():
     """Mock RetrieverGraph for streaming tests."""
     # Reset the mock for each test
     mock_graph_global.reset_mock()
+    # Set up the lazy-loaded graph so get_graph() returns the mock
+    conversations._rg = mock_rg_instance
+    conversations._rg_started.set()
+    conversations._rg_ready.set()
     yield mock_graph_global
+    # Teardown: reset graph state for next test
+    conversations.reset_graph_state_for_testing()
 
 
 @pytest.fixture
@@ -436,16 +442,15 @@ async def test_get_response_stream_graph_not_initialized(
         """Test behavior when graph is not initialized."""
         from src.api.routers.conversations import get_response_stream
 
-        with patch("src.api.routers.conversations.rg") as mock_rg:
-            mock_rg.graph = None
+        # Reset to simulate a fresh process where the graph hasn't been started yet
+        conversations.reset_graph_state_for_testing()
 
-            chunks = []
-            async for chunk in get_response_stream(sample_user_input, db_session):
-                chunks.append(chunk)
+        chunks = []
+        async for chunk in get_response_stream(sample_user_input, db_session):
+            chunks.append(chunk)
 
-            # When graph is None, streaming continues but produces no content chunks
-            # Should still have sources line
-            assert any("Sources:" in c for c in chunks)
+        # When graph is None, the stream yields an error and returns early
+        assert any("still initializing" in c for c in chunks)
 
     @pytest.mark.asyncio
     async def test_get_response_stream_empty_content(

docker-compose.yml

@@ -45,7 +45,7 @@ services:
       interval: ${HEALTHCHECK_INTERVAL:-30s}
       timeout: ${HEALTHCHECK_TIMEOUT:-10s}
       retries: ${HEALTHCHECK_RETRIES:-5}
-      start_period: ${HEALTHCHECK_START_PERIOD:-1200s}
+      start_period: ${HEALTHCHECK_START_PERIOD:-30s}
 
   frontend:
     build: