Query Rewriting
An optimization technique used in RAG and information retrieval systems that transforms natural language queries from users or agents into formats more suitable for search engine processing, with the goal of improving recall rates and retrieval relevance.
Timeline (last 90 days)
The difference between Query Fan-out and traditional query rewriting and query expansion is that the latter two generally still revolve around a single retrieval intent, whereas fan-out breaks a question into multiple mutually independent sub-queries executed separately, retrieving from different data sources and then integrating the results.
Query Rewriting solves the problem of users' questions being colloquial and having unclear references, and common methods include using a large model to rewrite the question into a formulation more conducive to retrieval, expanding a single-turn question into multiple related sub-questions, and completing the context based on historical conversation.
A typical approach to query rewriting is to design a dedicated prompt template that passes the last N rounds of conversation history together with the current user input to an LLM, instructing it to output a semantically complete, independently retrievable query statement.
The core of multi-turn dialogue query rewriting is to use large language models to understand and compress the context of multi-turn dialogues, completing vague follow-up questions into complete, self-contained query statements.
Advanced implementations of query rewriting include intent recognition, coreference resolution, and multi-intent decomposition.
All Facts (5)
The difference between Query Fan-out and traditional query rewriting and query expansion is that the latter two generally still revolve around a single retrieval intent, whereas fan-out breaks a question into multiple mutually independent sub-queries executed separately, retrieving from different data sources and then integrating the results.
50%UnverifiedQuery Rewriting solves the problem of users' questions being colloquial and having unclear references, and common methods include using a large model to rewrite the question into a formulation more conducive to retrieval, expanding a single-turn question into multiple related sub-questions, and completing the context based on historical conversation.
50%UnverifiedA typical approach to query rewriting is to design a dedicated prompt template that passes the last N rounds of conversation history together with the current user input to an LLM, instructing it to output a semantically complete, independently retrievable query statement.
50%UnverifiedThe core of multi-turn dialogue query rewriting is to use large language models to understand and compress the context of multi-turn dialogues, completing vague follow-up questions into complete, self-contained query statements.
50%UnverifiedAdvanced implementations of query rewriting include intent recognition, coreference resolution, and multi-intent decomposition.
50%