Truncated output

[berdeter]

Hello

I've developed a simple RAG example with Guardrails based on some content extracted from Wikipedia. It does work but I face a problem and find no solution. Results are truncated.

from nemoguardrails import LLMRails, RailsConfig
from langchain.chat_models import ChatOpenAI

# Lisons les fichiers de configuration
with open("astro.co", encoding="utf-8") as f:
    colang = f.read()

config = RailsConfig.from_content(colang_content=colang)


# Initialisons les rails
rag_rails = LLMRails(config, llm=ChatOpenAI(max_tokens=3000))

rag_rails.register_action(action=retrieve, name="retrieve")
rag_rails.register_action(action=rag, name="rag")

rag_rails.generate(prompt="Quelle est la différence entre une étoile à neutrons et un magnétar ?")

RAG Called
' La principale différence entre une étoile à neutrons et un magnétar est que les magnétars possèdent un champ magnétique extrêmement intense et émettent un rayonnement électromagnétique de haute énergie (rayons X et rayons gamma). Les étoiles à neutrons possèdent également un champ magnétique, mais il est moins intense et ell'

You see in the result it sundenly truncates content in the middle of a word ("elle" truncated to "ell")

As you see, I've already tried to move to GPT-35 instead of text-davinci and set max_tokens to 3000. It does not seem to affect the result that keeps being truncated after the same length.

Any idea is welcome...

Bernard