Intent Classification

Maps user queries to a predefined ontology of intents, such as purchase_inquiry, technical_support, or definition_lookup. This structured taxonomy enables deterministic routing to specialized downstream handlers. Key design considerations include:

• Granularity balance: Too fine-grained leads to classifier confusion; too coarse loses routing precision
• Mutual exclusivity: Intent boundaries must be clearly delineated to prevent ambiguous multi-intent matches
• Hierarchical nesting: Intents like billing.dispute.chargeback allow for fallback routing when confidence is low at leaf nodes

Modern intent classifiers leverage large language models to recognize intents with minimal or no training examples. A zero-shot classifier can categorize a query like "Why was my card declined?" into payment_failure without ever seeing that exact phrasing, using only a natural language description of each intent. Few-shot approaches provide 3-5 examples per intent to anchor model behavior, achieving production-grade accuracy for domains with rapidly evolving taxonomies.

Every classification decision carries an associated confidence score that gates downstream action. Queries falling below a configurable threshold—typically 0.7 to 0.85—are routed to a fallback strategy:

• Clarification prompts: The system asks the user to disambiguate between top-k candidate intents
• Default catch-all: Routes to a general-purpose retrieval pipeline when no intent is confidently identified
• Human escalation: High-stakes domains like healthcare or finance may flag low-confidence classifications for agent review

Real-world queries often contain compound goals that require decomposition. A query like "Compare the battery life and camera specs of the Pixel 9 and iPhone 16" contains both a product_comparison intent and multiple specification_lookup sub-intents. Multi-label classification architectures output a vector of binary decisions rather than a single argmax, enabling parallel retrieval across multiple intent-specific indexes.

Classified intents directly parameterize the retrieval pipeline by injecting structured filters into the search query. An intent of recent_news automatically appends a date range filter for the last 7 days. An intent of internal_policy restricts the search corpus to a specific document collection with appropriate access controls. This tight coupling between intent and retrieval logic eliminates the need for the user to manually specify search constraints.

User behavior and vocabulary evolve over time, causing intent drift where the distribution of queries shifts away from the original training data. Production intent classification systems employ:

• Confidence distribution tracking: Alerts when the mean confidence score degrades over a rolling window
• Novelty detection: Identifies clusters of queries that don't map well to any existing intent, signaling the need for taxonomy expansion
• Human-in-the-loop relabeling: Periodically samples low-confidence predictions for expert review and model retraining

The process of transforming a raw user query into one or more alternative formulations to improve retrieval recall. Once an intent classifier identifies the query's goal, rewriting expands or refines the input to bridge the lexical gap between user language and indexed documents.

• Expansion: Adding synonyms or related terms (e.g., 'CPU' → 'central processing unit')
• Decomposition: Breaking a complex query into sub-queries for multi-hop retrieval
• Normalization: Standardizing entities and resolving abbreviations

Effective rewriting is often conditioned on the predicted intent, applying domain-specific rules for each category.

A pre- or post-retrieval mechanism that restricts the candidate document set based on structured attributes. Intent classification directly maps user goals to specific filter parameters, ensuring results meet precise non-semantic constraints.

• Pre-Filtering: Apply constraints before vector search to guarantee all results match
• Post-Filtering: Apply constraints after ANN search, risking empty result sets
• Intent-Driven Examples:
  • 'Latest docs' intent → filter date > now() - 30d
  • 'Internal policies' intent → filter access_level = 'employee'
  • 'Product specs' intent → filter doc_type = 'datasheet'

A strategy that combines results from multiple heterogeneous retrieval systems to improve overall recall and robustness. The intent classifier acts as the gating mechanism, selecting which retrieval paths to activate and how to weight their contributions.

• Dense Vector Search: For semantic, conceptual queries
• Sparse BM25: For exact keyword and entity matching
• Knowledge Graph Traversal: For structured, relational queries
• Hybrid Example: A 'troubleshooting' intent might weight dense retrieval higher, while a 'compliance check' intent might prioritize exact keyword matching via BM25

A re-ranking methodology where a transformer model processes the concatenated query and candidate document simultaneously to produce a fine-grained relevance score. Intent classification can select different reranking models optimized for specific query types.

• Precision vs. Latency: Cross-encoders offer higher precision than bi-encoders but are computationally expensive
• Intent-Specific Models:
  • Factual queries → reranker trained on NLI datasets
  • Procedural queries → reranker trained on how-to corpora
  • Comparative queries → reranker sensitive to contrastive language
• Pipeline Position: Typically applied to top-k candidates from first-pass retrieval

A cascading pipeline architecture where an initial, computationally cheap first-pass retrieval stage generates a broad set of candidates, which is then progressively refined by more expensive and precise downstream stages. Intent classification determines the configuration of each stage.

• Stage 1 (Candidate Generation): Fast ANN or inverted index lookup, recall-focused
• Stage 2 (Re-ranking): Cross-encoder or feature-based scoring on top-1000 candidates
• Stage 3 (Calibration): Score normalization and thresholding based on intent confidence
• Dynamic Routing: Different intent classes can skip stages or use entirely different pipeline configurations to optimize for latency or precision