Contextual Query Expansion

The system analyzes the dialogue state and preceding turns to extract salient entities and intents. Unlike static synonym expansion, this method identifies latent conceptual links—for example, resolving that 'it' in a follow-up question refers to a specific product mentioned three turns earlier. This process relies on coreference resolution and entity resolution to build a coherent context graph before query reformulation.

Not all conversational context is equally relevant. Expansion algorithms apply recency-based weighting functions to historical tokens, giving higher salience to the immediate exchange while progressively decaying the influence of older utterances. This prevents context pollution where stale information from the beginning of a long session distorts the current retrieval intent.

The expanded query is executed against both sparse inverted indexes (BM25) and dense vector stores simultaneously. Contextual terms are converted into multiple representations:

• Lexical expansion: Adding synonyms and morphological variants for keyword matching
• Semantic expansion: Generating embedding vectors for the enriched query to capture conceptual neighbors This dual approach ensures high recall without sacrificing precision.

Expansion is gated by the classified user intent. For a transactional query ('buy a laptop'), the system injects product attributes and comparison terms. For an informational query ('explain the error'), it injects troubleshooting and diagnostic terminology. This intent-aware filtering prevents irrelevant term injection that would otherwise introduce noise into the retrieval pipeline.

If initial retrieval yields low-confidence results, the system enters a reformulation loop:

• The original query is re-expanded using alternative context interpretations
• A cross-encoder reranker evaluates candidate documents against the full conversation
• Failed expansions are logged to refine future term selection models This iterative process mirrors the ReAct framework of reasoning and acting.

Ambiguous terms like 'apple' are disambiguated by analyzing the conversational context window for co-occurring entities. If the dialogue mentions 'iPhone' and 'iOS', the expansion injects 'Apple Inc.' and technology-related terms. If the context includes 'orchard' and 'harvest', it injects agricultural terminology. This contextual entity linking prevents catastrophic semantic drift in the expanded query.