Query Intent Classification

The user seeks knowledge or an answer to a question. These queries often begin with 'what is,' 'how to,' or 'why does.' The goal is to retrieve comprehensive, authoritative content.

• Retrieval Strategy: Heavily weight dense vector search to capture semantic meaning and context.
• Fusion Tuning: Apply semantic boosting to prioritize conceptual matches over exact keywords.
• Example: 'How does reciprocal rank fusion combine sparse and dense results?'

The user wants to locate a specific website, page, or known item. The query often contains brand names, product codes, or exact titles.

• Retrieval Strategy: Heavily weight sparse lexical matching (BM25) for precise term matching.
• Fusion Tuning: Apply lexical boosting to ensure exact matches on identifiers like 'SKU-4523' rank first.
• Example: 'Elasticsearch hybrid search documentation'

The user intends to perform an action, such as downloading, purchasing, or subscribing. Queries contain action verbs like 'buy,' 'download,' or 'register.'

• Retrieval Strategy: Combine semantic understanding with metadata filtering to surface actionable pages.
• Fusion Tuning: Use pre-filtering on metadata (e.g., product availability) before semantic re-ranking.
• Example: 'Download latest white paper on agentic observability'

A subset of transactional intent where the user is comparing products or services before a decision. Queries include terms like 'best,' 'vs,' 'review,' or 'top 10.'

• Retrieval Strategy: Fuse dense retrieval for topical breadth with a Cross-Encoder Re-Ranker for precision on comparative content.
• Fusion Tuning: Apply Weighted Sum Fusion to balance semantic relevance with recency signals.
• Example: 'ColBERT vs Cross-Encoder re-ranking latency comparison'

When intent classification confidence is low, the system activates a predefined fallback strategy to prevent empty results. This often reverts to pure lexical search or applies aggressive query relaxation.

• Trigger: Ambiguous one-word queries or high entropy in the classifier's softmax output.
• Action: Default to Reciprocal Rank Fusion (RRF) with equal weights, bypassing intent-specific boosting.
• Example: A query containing only 'fusion' with no contextual modifiers.

The classified intent dynamically adjusts the fusion weight assigned to sparse and dense retrieval scores before merging. This is a core optimization in modern Hybrid Search architectures.

• Informational: Dense weight = 0.8, Sparse weight = 0.2
• Navigational: Dense weight = 0.2, Sparse weight = 0.8
• Transactional: Dense weight = 0.5, Sparse weight = 0.5 with strict metadata pre-filtering
• Optimization: Weights are often tuned via Bayesian optimization against NDCG@10.

A query type where the user aims to reach a specific website or known destination. The system identifies these queries to bypass heavy semantic re-ranking and route directly to the authoritative result.

• Example: "inferensys docs hybrid search"
• Action: Boosts Lexical Boosting and exact match signals
• Fusion Impact: Reduces dense vector weight to prevent semantic drift from the target URL

A query type seeking broad knowledge or answers to open-ended questions. Classification triggers deep semantic retrieval and multi-document synthesis.

• Example: "how does reciprocal rank fusion work"
• Action: Activates Semantic Boosting and dense passage retrieval
• Fusion Impact: Increases vector search weight to capture conceptual matches beyond exact keywords

A query indicating the user intends to perform a specific action, such as downloading, purchasing, or executing an API call. Classification routes these to structured action endpoints.

• Example: "download latest whitepaper on agentic memory"
• Action: Triggers Tool Calling and API execution pipelines
• Fusion Impact: May bypass traditional search indexes entirely in favor of direct function invocation

The direct consequence of intent classification in a hybrid search architecture. The predicted intent class dynamically adjusts the Weighted Sum Fusion parameters at query time.

• Navigational: Weight heavily favors sparse BM25 scores
• Informational: Weight shifts toward dense Bi-Encoder Scoring
• Transactional: System may override fusion entirely for a Fallback Strategy to an action handler

A preprocessing step often coupled with intent classification. Before intent is determined, raw queries are normalized to improve classification accuracy and downstream retrieval.

• Techniques: Acronym expansion, spelling correction, entity linking
• Synergy: A query rewritten from "RRF meaning" to "Reciprocal Rank Fusion definition" is more reliably classified as Informational Intent
• Pipeline Order: Normalization occurs before the intent classifier

A safety mechanism triggered when intent classification confidence is low or the primary retrieval path returns zero results. The system reverts to a robust default behavior.

• Low Confidence: Defaults to a balanced Sparse-Dense Hybrid fusion
• Zero Results: Expands the query via Query Expansion Techniques and retries
• Goal: Prevents empty result pages and ensures graceful degradation