To calculate the final ranking score for the top 10 results using complex features.

To achieve the highest possible precision by deeply analyzing a small set of documents.

To understand user intent and expand the query with synonyms.

To maximize recall by quickly retrieving a large set of potentially relevant documents from the entire collection.

It requires an inverted index, whereas TF-IDF does not.

It incorporates pre-trained language model embeddings.

It accounts for term frequency saturation and document length normalization.

It exclusively uses boolean 'AND' logic for all query terms.

Inverted Index

BM25 Algorithm

A/B Testing Framework

Dense Retrieval using Bi-Encoders

To store the original text of the documents associated with each embedding.

To fine-tune the BERT encoder for better domain-specific embeddings.

To provide a highly efficient, sub-linear time method for finding the closest vectors to a query vector, at the cost of perfect accuracy.

To calculate the final NDCG score for a ranked list of results.

To categorize the user's underlying goal (e.g., informational, navigational, transactional) to tailor the search results page.

To correct spelling mistakes and expand abbreviations.

To identify specific products, people, or locations mentioned in the query.

Pointwise

Listwise

Pairwise

True

False

Normalized Discounted Cumulative Gain (NDCG)

Click-Through Rate (CTR)

Mean Average Precision (MAP)

Mean Reciprocal Rank (MRR)

To perform A/B testing between a neural model and a lexical baseline.

To manage the latency-precision trade-off by using progressively slower, more accurate models at each stage of ranking.

To fine-tune a single large transformer model on multiple search-related tasks simultaneously.

To chain multiple query understanding models together.

The historical click-through rate of the user.

The number of terms in the user's query.

The BM25 score between the query and the document.

The PageRank score of the document.

To normalize the scores from different retrieval systems (e.g., sparse BM25 and dense vector search) before combining them.

To increase the speed of the dense vector search component.

To replace the need for a Learning to Rank model.

To exclusively select documents that appear in the results of all retrieval systems.

Query Expansion using Pseudo-Relevance Feedback

Intent Classification

Entity Extraction

HNSW Indexing

NDCG Calculation

Counterfactual Evaluation (e.g., Inverse Propensity Scoring)

A/B Testing

Dot product is computationally more expensive than cosine similarity.

Dot product cannot be used in vector databases.

Cosine similarity is bounded between -1 and 1 and is invariant to vector magnitude, focusing solely on orientation.

Cosine similarity works for sparse vectors while dot product only works for dense vectors.