The parser decides which failures you see.

Eight replies a model could give when asked for a JSON answer with citations, run through four LangChain output parsers. The replies are hand-written and representative; no model was called. The parsers are ports of langchain-core 1.7.0, checked against the library.

Output parser at the end of the chain

Returned without error
Raised OutputParserException
Returned, but unusable
Usable answers

What it adds to the prompt

Show the format instructions

Eight replies, one parser

Question: “How long does a refund take?” Passages: refund-window, refund-timing, billing.

Reply (select to inspect)	Result	What your code receives

Usable means an answer string and at least one citation that names a real passage. The parsers do not know that rule; the lab grades every result against it.

Inspect one reply

Model reply for ; edit it to test your own

Cut the reply after this many characters, as a max_tokens limit would

Parser	Result	Detail

Every cut of this reply	Raised	Unusable	Usable