Query rewriting¶

QueryRewriting is a small, experimental helper in luckybrain.query_imp. It asks a chat model to rephrase search text. It does not retrieve data, validate the model's factual accuracy, or guarantee semantic equivalence.

Usage¶

Pass an object that implements the synchronous think(prompt: str) -> str method, such as either built-in chat client:

from luckybrain.core.models import OpenaiBrain
from luckybrain.query_imp import QueryRewriting

rewriter = QueryRewriting(
    model=OpenaiBrain(model="gpt-4o-mini"),
    style="search",
)

clean_query = rewriter.improve("  how do i reset password?? ")
alternatives = rewriter.variants("reset account password", n=3)

Constructor¶

Parameter	Type	Default	Meaning
`model`	object with `think(prompt: str) -> str`	Required	Synchronous model used to generate rewrites.
`style`	`str \\| None`	`"search"`	Exactly `"search"` requests keyword-focused wording. Any other value requests a well-formed question.

Methods¶

Method	Input	Result
`improve(query)`	A string. Whitespace-only input returns an empty string without a model call.	One rewritten string, stripped of surrounding whitespace and quote characters.
`variants(query, n=3)`	Search text and an optional result cap.	Up to `n` non-empty lines from the model, with common bullet prefixes removed.

The number and quality of returned variants are not guaranteed. The helper does not validate the model output or retry when fewer than n variants are returned. The model controls cost, latency, privacy handling, and failure behavior; its exceptions propagate to the caller.

When to use it¶

Use query rewriting as an optional step when broadening or cleaning a search query is useful. Preserve the original query as a fallback, and evaluate retrieval quality with your own data before relying on rewrites. Do not use generated alternatives as evidence or silently discard the user's intent.

This API is experimental and is not re-exported from the root luckybrain package. Import it from luckybrain.query_imp.