LangChain

pip install "llm-beacon[langchain]"

from langchain_openai import ChatOpenAI
from pydantic import BaseModel
from beacon.integrations.langchain import BeaconJsonOutputParser, BeaconPromptTemplate

class Intent(BaseModel):
    intent: str
    confidence: float

prompt = BeaconPromptTemplate.from_template("Classify: <msg>{message}</msg>")
parser = BeaconJsonOutputParser(pydantic_object=Intent)

chain = prompt | ChatOpenAI(model="gpt-4o") | parser
chain.invoke({"message": "I want to cancel"})   # -> Intent(intent='cancel', confidence=0.93)

`BeaconPromptTemplate`

A StringPromptTemplate rendered by template:

input variables are inferred from the template;
values are XML-escaped (escape= to change);
missing variables raise MissingVariableError;
extra chain inputs are ignored, as LangChain expects.

Format instructions contain JSON braces, so pass them as a partial and insert them unescaped:

prompt = BeaconPromptTemplate(
    template="<question>{question}</question>\n\n{format_instructions!raw}",
    partial_variables={"format_instructions": parser.get_format_instructions()},
)

`BeaconJsonOutputParser`

Strict extraction plus schema validation. Pass pydantic_object= (returns model instances), json_schema=, or pydantic_schema=Model.model_json_schema().

Failures raise LangChain’s OutputParserException, so with_retry() and fixing parsers keep working. The beacon error is attached as .beacon_error and as __cause__:

try:
    chain.invoke(...)
except OutputParserException as exc:
    exc.beacon_error.code     # 'json.trailing_comma'
    exc.llm_output            # the raw model text

get_format_instructions() is deterministic for a given schema.

`BeaconCallbackHandler`

Records what beacon makes of every LLM response in a run, without changing the chain:

handler = BeaconCallbackHandler(schema=Intent.model_json_schema())
chain.invoke(inputs, config={"callbacks": [handler]})
handler.records[-1]
# {'run_id': '…', 'text_hash': '1593ddc6f67bfc10', 'json_found': True,
#  'source': 'fence', 'valid': True, 'errors': [], 'error_code': None}

Records are also logged at INFO on the beacon.integrations.langchain logger. They contain only deterministic data (no timestamps), so you can snapshot them.