[Python] How can I reduce the startup time of FastAPI

I'm building a Retrieval-Augmented Generation (RAG) application using Python + FastAPI. I'm also running a Kafka consumer alongside the FastAPI app to process incoming text and generate embeddings (stored in FAISS). However, I'm facing a major issue — the startup time for both the Kafka worker and FastAPI app is very high (90–100 seconds). I tried optimizing it using FastAPI’s lifespan event to offload warm-up tasks (model loading, vector store loading, etc.) into the background, which did reduce it slightly to ~85–88 seconds — but it's still too slow for practical use. Here's my question:

what should i do to reduce the startup time of my kafka worker and fastapi app

here's the code snippet for reference:

I'm trying to defer heavy initialization during FastAPI startup using lifespan like this:

@asynccontextmanager
async def lifespan(app: FastAPI):
logger.info("Starting up RecallAI...")
start = time.time()

async def warmup():
logger.info("Warming up models and vector store...")
get_embeddings_model()
get_llm()
get_vectorstore()
logger.info(f"Warmup complete in {time.time() - start:.2f} seconds")

asyncio.create_task(warmup())
yield
logger.info("Shutting down RecallAI...")


This helps, but not enough — the app still takes a long time to become responsive.

I think the possible reasons are:

1. 
   Most of the startup delay seems to come from store_embeddings() and model/vectorstore loading.
   
   
2. 
   Even importing store_embeddings in the Kafka worker takes multiple seconds.
   
   
3. 
   I'm using langchain, FAISS, and HuggingFace for embeddings.
   

FastAPI app + lifespan logic:

@cache
def get_embeddings_model():
return HuggingFaceEmbeddings(model_name="BAAI/bge-small-en-v1.5")

@cache
def get_llm():
return ChatGroq(groq_api_key=os.getenv("GROQ_API_KEY"), model_name="llama-3-70b")

def get_vectorstore():
embeddings = get_embeddings_model()
return FAISS.load_local("img_vector_store", embeddings, allow_dangerous_deserialization=True)


Here's the Kafka worker:

from recall_ai.vector_embeddings.store_vector_embedding import store_embeddings

def insert_batch(batch):
try:
res = store_embeddings()
...


Even importing store_embeddings takes 90 seconds. This function internally loads models and creates FAISS indexes, which contributes to the delay.

Kafka worker still taking upto 90sec to start

Continue reading...