Prompt versions

create takes label or version. Without both, the factory label applies, and without that one the production label.

staging = LangfuseChainFactory(langfuse, models=models, label="staging")

chain = staging.create("support-reply")  # label "staging"
live = staging.create("support-reply", label="production")
pinned = staging.create("support-reply", version=12)  # the factory label does not apply

Every other keyword argument of create, such as cache_ttl_seconds, max_retries or fetch_timeout_seconds, is passed to client.get_prompt as it is:

chain = factory.create("support-reply", cache_ttl_seconds=30)

Cache and replicas

The SDK caches each prompt and refreshes it in the background after cache_ttl_seconds, 60 by default. A label move reaches a running service within that time, and until it does, replicas may run different versions. cache_ttl_seconds=0 fetches on every call.

Warming up

The first call of a chain fetches the prompt over HTTP and builds template | model. Concurrent first calls fetch once. To take that off the first request, warm the chain up at startup:

chain.get_input_schema()  # fetches the prompt and builds the chain

A cold start while Langfuse is unreachable raises PromptFetchError; see Error handling for a local fallback.