LlamaIndex

pip install "llm-beacon[llamaindex]"

from llama_index.core import SimpleDirectoryReader, VectorStoreIndex
from llama_index.core.output_parsers import PydanticOutputParser
from pydantic import BaseModel
from beacon.integrations.llamaindex import BeaconOutputParser

class QueryResponse(BaseModel):
    answer: str
    sources: list[str]
    confidence: float

index = VectorStoreIndex.from_documents(SimpleDirectoryReader("data").load_data())
parser = BeaconOutputParser(PydanticOutputParser(QueryResponse))
query_engine = index.as_query_engine(output_parser=parser)

The parser sits between LlamaIndex’s response synthesizer and your code:

• format() delegates to the wrapped parser, so LlamaIndex’s format instructions still reach the model.
• parse() uses beacon’s strict extraction and validation instead of lenient JSON scraping, and returns a QueryResponse.

BeaconOutputParser subclasses LlamaIndex’s BaseOutputParser. Construct it in any of these ways:

BeaconOutputParser(PydanticOutputParser(QueryResponse))   # reuse output_cls + format()
BeaconOutputParser(output_cls=QueryResponse)
BeaconOutputParser(json_schema={"type": "object", "required": ["answer"]})   # returns dicts

Failures raise MalformedOutputError or SchemaValidationError with .raw set to the synthesized text:

>>> parser.parse('{"answer": "30 days", "sources": "refunds.md", "confidence": 0.86}')
SchemaValidationError: 1 schema violation:
  - /sources: expected type "array", got "string"