Conversational Query Reformulation

The foundational NLP task of identifying all expressions that refer to the same entity. In a dialogue, this resolves anaphora by linking pronouns and definite noun phrases to their antecedents.

• Example: In "Show me the Q3 report. It was finalized yesterday," the system resolves it to "the Q3 report."
• Mechanism: Modern systems use span-based neural models that score candidate antecedent spans within the context window.
• Impact: Without this step, the reformulated query would be semantically incomplete, leading to retrieval failure.

The process of reconstructing omitted words or phrases that are syntactically required but contextually implied. Ellipsis is common in multi-turn dialogues where users avoid repetition.

• Example: User 1: "What is the price of the enterprise plan?" User 2: "And the pro plan?" The system must expand the second query to "What is the price of the pro plan?"
• Approach: Sequence-to-sequence models are trained on dialogue corpora to predict the complete utterance from the partial input and conversation history.

The strategy for selecting and truncating dialogue history to provide as input for reformulation. The system must balance informativeness with the token limits of the reformulation model.

• Recency Bias: Prioritizing the most recent turns, as they are most likely to contain the relevant antecedent.
• Entity Tracking: Maintaining a running set of active entities across the session to ensure long-range dependencies are not lost.
• Sliding Window: Using a fixed number of past turns, with older turns dropped when the limit is reached.

The core technique of using a fine-tuned LLM to rewrite the current user utterance into a standalone query. The model is prompted with the conversation history and instructed to produce a self-contained string.

• Training Data: Pairs of context-dependent utterances and their manually de-contextualized forms.
• Zero-Shot Prompting: Instruction-tuned models can perform this task without fine-tuning by following a directive like "Rewrite the following query to be fully self-contained, incorporating necessary context from the history."
• Output: A single, unambiguous query string ready for the retrieval pipeline.

A critical constraint ensuring the reformulated query does not alter the user's original objective. The transformation must add context without shifting the underlying information need.

• Validation: Comparing the intent classification of the original utterance and the reformulated query to detect drift.
• Negative Example: User: "What about France?" (in a discussion about sales regions). A failure would be reformulating to "Tell me about France," which is too broad. A success would be "What were the Q3 sales figures for the France region?"

The mechanism for anchoring queries to implicit temporal references and session-specific filters that persist across turns.

• Temporal Anchoring: Resolving "last month" or "yesterday" to absolute date ranges based on the current timestamp.
• Filter Persistence: If a user previously filtered results to a specific department, the reformulated query must carry that constraint forward unless explicitly cleared.
• Implementation: A state object is maintained alongside the dialogue, updated with each turn, and injected into the reformulation prompt.