Hybrid Search

Uses embedding models to map queries and documents into a high-dimensional vector space where semantic similarity is measured by cosine similarity. This component excels at understanding conceptual relationships, synonyms, and paraphrased content that keyword matching would miss.

• Captures latent semantic meaning beyond exact term overlap
• Handles natural language queries with high variability
• Requires an ANN index like HNSW for efficient similarity search at scale

Employs the BM25 probabilistic ranking function to perform exact term matching with TF-IDF weighting. This component provides precision for rare, domain-specific terms like product codes, legal citations, or technical identifiers that embedding models may overlook.

• Guarantees exact match for critical keywords and entity names
• Excels at matching alphanumeric codes and specialized jargon
• Computationally lightweight compared to dense vector search

A score-agnostic fusion algorithm that combines ranked result lists from dense and sparse retrievers without requiring calibration of their raw scores. RRF assigns a reciprocal score to each document based on its rank position across both lists, producing a unified ranking.

• Formula: score(d) = Σ 1/(k + rank_i(d)) where k is typically 60
• Eliminates the need for score normalization between heterogeneous retrievers
• Robust to outliers and scale differences in underlying scoring functions

An alternative fusion strategy that normalizes the raw similarity scores from dense and sparse retrievers and combines them using a weighted linear formula. This approach allows fine-tuning the balance between semantic and keyword influence.

• Formula: final_score = α × dense_score + (1-α) × sparse_score
• The alpha parameter controls the semantic vs. keyword trade-off
• Requires careful score normalization to prevent one retriever from dominating

Combines hybrid retrieval with metadata extraction to apply pre-filtering or post-filtering based on structured attributes like date ranges, document types, or access control labels. This ensures results are not only semantically and lexically relevant but also contextually valid.

• Pre-filtering narrows the candidate set before vector search
• Post-filtering removes invalid results after ranking
• Critical for enforcing document-level security permissions in enterprise deployments

A multi-stage architecture where hybrid retrieval generates a coarse candidate set, and a cross-encoder model performs computationally intensive joint scoring of each query-candidate pair to produce a precision-optimized final ranking.

• Stage 1: Fast hybrid retrieval retrieves top-N candidates (e.g., N=100)
• Stage 2: Cross-encoder re-ranks candidates with full attention computation
• Delivers the recall of hybrid search with the precision of deep interaction models

A hyperparameter tuning strategy that controls the relative influence of sparse and dense retrieval scores during fusion. The weighting coefficient α determines the balance:

• α = 0.0: Pure dense vector search
• α = 1.0: Pure BM25 keyword search
• α = 0.5: Equal contribution from both

Optimal α values are domain-dependent and should be tuned using evaluation metrics like NDCG@10 on representative query sets. Some systems dynamically adjust α per query based on intent classification.

A preprocessing step that categorizes incoming queries to determine the optimal retrieval strategy. Intent types directly influence hybrid search behavior:

• Exact lookup: Heavily weight BM25 (e.g., error code searches)
• Conceptual exploration: Heavily weight dense retrieval (e.g., 'best practices for scaling')
• Mixed intent: Use balanced fusion weights

Intent classifiers are typically fine-tuned distilBERT or T5 models trained on query logs with human-annotated categories.