Query Reformulation

The core mechanism distinguishing reformulation from static expansion. The system analyzes the initial retrieval set or explicit user signals to identify a mismatch between intent and results. It then generates a modified query, executes it, and repeats the process. This loop continues until a termination condition is met, such as result stability, a confidence threshold, or user satisfaction. This is the foundation of relevance feedback systems.

Reformulation leverages the full search session history, not just the current query string. It models the user's evolving information need by analyzing previous queries, clicked documents, and dwell time. For example, a follow-up query of 'How do I deploy it?' is meaningless without resolving the anaphora 'it' to the entity from the prior turn (e.g., 'Docker container'). This requires robust coreference resolution and short-term session state management.

A critical challenge is balancing refinement with concept drift. Reformulation must clarify ambiguous intent without straying from the original goal. Techniques include:

• Query Relaxation: Removing overly restrictive terms that caused zero results.
• Query Scoping: Adding categorical filters when results are too broad.
• Term Weighting: Boosting or dampening specific concepts based on their diagnostic value in the initial results. The system must distinguish between a poorly formulated query and a genuinely novel information need.

An automatic reformulation technique that operates without explicit user input. PRF assumes the top-k documents from an initial retrieval are relevant. It then extracts the most statistically significant terms from these pseudo-relevant documents and appends them to the original query for a second retrieval pass. While powerful for boosting recall, its performance is brittle and can catastrophically fail if the initial top-k documents are non-relevant, a problem known as query drift.

In dense retrieval systems, reformulation occurs in the embedding space. The Rocchio algorithm is a classic example: it reformulates a query vector by moving it closer to the centroid of relevant document vectors and away from the centroid of non-relevant ones. Modern neural approaches use query2vec models or fine-tuned encoders that take the initial query and a feedback document set as input to generate a new, refined query embedding directly.

Large Language Models (LLMs) enable sophisticated, natural language reformulation. Given the original query and a summary of the initial poor results, an LLM can be prompted to generate multiple alternative queries that disambiguate intent, fill lexical gaps, or break a complex question into simpler sub-queries. This moves beyond term-based manipulation to true paraphrase generation and conceptual restructuring, often using chain-of-thought reasoning to hypothesize why the first query failed.