Search and Retrieval

Prime #

160

Origin domain

Computer Science & Software Engineering

Also from

Library Information Science, Cognitive Science, Biology & Ecology

Aliases

Information Retrieval, Query Resolution, Lookup, Search

Related primes

Index, Order, heuristic search, Pattern Recognition, database query

Core Idea

Search and Retrieval is the process of locating, identifying, and retrieving relevant information, resources, or objects from a larger dataset, environment, or memory system, often optimizing for speed, accuracy, and efficiency. The essential commitment is that given a query or information need, a system must navigate a search space (continuous or discrete, structured or unstructured) to discover items matching specified criteria, balancing exhaustiveness against computational cost^[1]. Every search-and-retrieval system faces trade-offs between precision (false positives excluded), recall (false negatives excluded), and query latency, and must determine both what is relevant and how efficiently to locate it.

How would you explain it like I'm…

Finding things

When you've lost your favorite toy, you check the toy box, then under the bed, then in the closet. You're searching: looking through places one by one until you find it. Computers do the same thing when you ask a question: they look through lots of stuff to find what matches.

Looking up stuff

Search and retrieval is about finding the thing you need inside a much bigger pile. Whether it's a word in a book, a video on the internet, or a memory in your brain, there's some space to look through and some idea of what counts as a match. Good searching balances two things: being fast and finding the right stuff. If you're too fast, you miss things; if you're too careful, it takes forever.

Locating matching items

Search and retrieval is the process of locating and pulling out relevant items from a larger collection, whether that's a database, the web, a library, or your own memory. Given a query (what you're looking for) and a search space (where to look), the system has to navigate that space and return the items that match. Every search system trades off three things: precision (how much of what it returns is actually relevant), recall (how much of the truly relevant stuff it manages to find), and speed (how long it takes). Different uses care about different trade-offs, a web search wants speed and precision, a legal discovery system wants recall above all.

 

Search and retrieval is the process of locating, identifying, and retrieving relevant items from a larger collection, dataset, or memory system in response to a query or information need. The system navigates a search space, which may be discrete or continuous, structured (a database with schemas) or unstructured (raw text, images), and returns items satisfying some relevance criterion. Every retrieval system faces three structural trade-offs: precision (the fraction of returned items that are relevant), recall (the fraction of relevant items that are returned), and latency (how long retrieval takes). These trade-offs interact with the size of the space, the richness of indexing, and the cost of computing relevance, so design choices, exact-match versus approximate, lexical versus semantic, exhaustive versus heuristic, depend on what the application most cares about.

Structural Signature

• The query or information need specification and formalization ^[1]
• The search space representation and traversal strategy (linear scan, tree index, hash index, graph traversal) ^[2]
• The relevance model and matching criterion (exact, fuzzy, semantic, ranked) ^[3]
• The indexing and pre-computation structures enabling sub-linear retrieval ^[2]
• The ranking and ordering function when multiple matches exist ^[4]
• The recall-precision-latency trade-off tuning at deployment time ^[5]

What It Is Not

• Not identical to sorting. Sorting orders a complete dataset; search retrieves a subset matching a criterion. A search may use sorting as a sub-operation but is not defined by it. You can search without sorting (hash lookup), and sort without searching (produce all elements in order).

• Not mere pattern matching. Pattern matching tests a single string or item against a pattern; search applies that test across many candidates and aggregates results. Pattern matching is a sub-operation; search is the larger orchestration.

• Not equivalent to filtering. Filtering removes items that do not meet criteria; search adds the dimension of how to locate candidates efficiently rather than inspecting all items. A filter applied naively to all data is inefficient search.

• Not a single algorithm. Linear search, binary search, hash-table lookup, B-tree traversal, inverted-index lookup, graph search, semantic vector similarity all solve search problems with different assumptions about data structure and query semantics.

• Not retrieval alone. Retrieval is the act of obtaining the item once found; search is the process of locating it. The two are often composed: search to identify, then retrieve to obtain.

• Common misclassification: Confusing "search" with "full-text search" (a specific instantiation), or assuming all search problems are solved by the same method regardless of data structure, index availability, or latency constraints.

Broad Use

Search and Retrieval appears across nearly every domain where information or resources must be located at scale. In operating systems, file system lookup uses inode tables and B-trees to traverse directory hierarchies; process scheduling queues are indexed by priority or deadline to efficiently select the next runnable task; memory management systems use page tables and TLBs to translate virtual addresses to physical ones. In databases, query optimization engines select among multiple index strategies (B-tree range scans, hash-join lookups, index-intersection algorithms) based on cardinality statistics and cost models; execution engines iterate through result cursors one row at a time to avoid materializing entire result sets. In search engines, inverted indexes map each term to lists of documents, with compression and skip lists enabling million-document retrieval in milliseconds; PageRank and other ranking signals order results by authority and relevance; distributed index sharding (documents partitioned across machines) enables parallel retrieval. In information retrieval systems, the Salton-McGill vector space model and BM25 probabilistic framework provide mathematically grounded relevance models; TREC (Text REtrieval Conference) benchmarks evaluate precision-recall trade-offs across systems. In machine learning, nearest-neighbor search in high-dimensional vector spaces (embeddings) enables recommendation and similarity tasks; locality-sensitive hashing (LSH) provides approximate retrieval in polynomial rather than exponential time. In memory systems, CPU caches implement set-associative lookup with LRU replacement; virtual memory uses multi-level page tables and TLBs for fast address translation. In cognitive psychology, memory retrieval is triggered by associative cues; priming effects show that recent or frequent memories surface faster. In biology, bacterial chemotaxis searches for nutrient gradients via tumble-and-run algorithms; predators forage by balancing search cost against capture success.

Clarity

Search and Retrieval clarifies that information access is not "free" — it requires a strategy. The construct separates the specification of what is wanted (query, information need) from the mechanism by which it is found (algorithm, index), making explicit the role of pre-computation (indexing, caching) in trading storage for retrieval speed. It forces recognition that relevance is defined rather than objective; the same query can return different results depending on the relevance model used^[3]. It shows that efficiency is not intrinsic to data but depends on access patterns and index design. The clarity also reveals why naive linear search (checking every item) fails at scale, motivating the design of indexes, caches, and approximate-retrieval methods. By making the query explicit, the construct enables reasoning about what constitutes a good answer: is precision critical (false positives are costly) or recall (false negatives are costly)? Does latency matter (interactive queries) or throughput (batch processing)? Can the system afford the storage overhead of sophisticated indexes, or must it operate on minimal resources?

Manages Complexity

The construct manages complexity by decomposing the search problem into layers: query parsing (what are we looking for?), index maintenance (how do we organize for rapid access?), search execution (how do we traverse the index?), and ranking (how do we order results?). Pre-computed indexes (inverted indexes, B-trees, hash tables, semantic vector embeddings) decouple query time from data size, moving expensive work to index build-time rather than query-time^[2]. This enables scaling from megabyte datasets to exabyte-scale search engines. Hierarchical and approximate retrieval methods (bounds-based pruning, locality-sensitive hashing, hierarchical navigable small-world graphs) reduce search-space exploration. The structural clarity also supports caching and memoization strategies — retrieving frequently-searched items from fast-path caches rather than full search. Finally, the framework enables reasoning about trade-offs: adding indexes speeds retrieval but increases storage and update cost; relaxing precision (accepting approximate matches) speeds retrieval and reduces index size. Query optimization decisions (which index to use, whether to parallelize) become explicit, analyzable, and tunable.

Abstract Reasoning

Search-and-retrieval reasoning proceeds by identifying the search space and its properties (size, structure, dimensionality, growth rate), specifying the relevance criterion or similarity metric (what does "relevant" mean for this domain?), selecting an index structure and retrieval algorithm suited to the workload (lookup-heavy vs. range-heavy vs. nearest-neighbor), and tuning the balance between recall, precision, and latency. It supports systems design (query optimizer decisions, cache eviction policies, hot-data identification), organizational decisions (what information to catalog and how, what metadata to maintain), cognitive strategies (which cues trigger memory retrieval most reliably), and biological fitness (time spent foraging vs. energy obtained)^[1]. Engineers ask: given expected query patterns (uniform random access vs. skewed popularity), what index structure minimizes latency while staying within memory constraints? Given a fixed computational budget, should we invest in preprocessing (index building) or runtime execution? If data is streaming and freshness matters, can we update indexes incrementally?

Knowledge Transfer

A software engineer's search-and-retrieval reasoning (query specification, index design, ranking, cache strategy) transfers across database query optimization, search-engine implementation, cache design, and memory-hierarchy tuning. The structural core is the insight that access patterns matter and can be optimized via pre-computation; what varies is the data structure (hash, tree, vector embedding, graph, bloom filter) and the relevance metric (exact equality, fuzzy string similarity, vector distance, semantic relevance). The same diagnostic framework — is the index appropriate for the query pattern, is precision/recall balanced correctly, is latency acceptable — applies to database indexes, full-text search, nearest-neighbor lookup, CPU caches, TLBs, and cognitive-memory cues. An engineer optimizing a web search engine uses the same principles as a neuroscientist studying memory recall.

Examples

Formal/abstract

The vector space model (Salton & McGill, 1983)^[1] represents documents and queries as vectors in a high-dimensional term space, where each dimension is a term (word) and the value is the term's frequency or importance (TF-IDF weighting). A query is likewise vectorized, and relevance is computed as the cosine similarity between the query vector and document vectors, with retrieval returning documents ranked by similarity. This formalism decouples the representation (vectors) from the retrieval method (cosine distance), enabling systematic study of ranking functions. Modern systems extend to semantic embeddings, where dense vectors capture meaning rather than term occurrence, enabling approximate nearest-neighbor search via locality-sensitive hashing or learned-index methods.

Mapped back: This instantiates the structural signature directly — query formalization (vector), search-space representation (high-dimensional space), relevance model (cosine similarity), ranking (similarity scores sorted), and efficiency (vector indexing structures).

Applied/industry

A web search engine indexes billions of documents using an inverted index: a mapping from each word to the list of documents containing it, with positions and frequency metadata. When a user queries "machine learning algorithms," the engine retrieves the posting lists for each term, intersects them to find documents containing all terms, applies ranking signals (PageRank, click history, recency, domain authority), and returns the top K results. The system uses distributed indexes (documents sharded across many machines), caching (popular queries cached with pre-computed result ranks), and approximate retrieval (early termination when confidence in top-K is high). Kubernetes' service discovery indexes services by name and labels, enabling efficient lookup of IP addresses for load balancing. Elasticsearch provides full-text search over inverted indexes with text analysis pipelines, filtering, faceting, and relevance tuning, powering analytics and log search systems.

Mapped back: These show search-and-retrieval as the unifying principle behind modern information systems, instantiated via indexing, ranking, and efficiency optimization in production environments.

Structural Tensions

• T1: Precision vs Recall vs Query Latency. Exhaustive search retrieves all relevant items (high recall) but is slow. Approximate or early-termination retrieval is fast (low latency) but misses some results (lower recall). Tightening relevance criteria improves precision (fewer false positives) but may reduce recall. No single point is optimal across all use cases^[5].

• T2: Index Maintenance Cost vs Retrieval Speed. Sophisticated indexes (B-trees, learned indexes, semantic embeddings) enable rapid retrieval but require expensive rebuild or incremental maintenance when data changes. Write-heavy systems struggle with index staleness; stale indexes reduce retrieval quality, but rebuilding is expensive. Real-time systems must balance: commit-log-based indexes trade some retrieval speed for fresh writes; batch indexing sacrifices freshness for efficiency.

• T3: Index Size and Memory Pressure. Larger indexes improve retrieval speed and recall but consume storage and RAM. Distributed indexes solve this by sharding but introduce network latency and coordination complexity. Systems with limited memory must prune indexes or use approximate methods, trading accuracy for size.

• T4: Query Complexity vs Expressiveness. Simple queries (keyword search, exact match) are fast to execute but inexpressive. Complex queries (boolean operators, faceted search, semantic constraints) are expressive but slow and difficult to optimize. The system must decide what query semantics to expose.

• T5: Centralized vs Distributed Search. A single search index is simple to manage and guarantees consistency but becomes a bottleneck at scale. Distributed indexes parallelize retrieval but introduce consistency, freshness, and coordination complexity^[6]. Geographic replication adds further trade-offs: local queries are fast but must handle cache-invalidation when indices diverge.

• T6: Semantic Relevance and Human Satisfaction. Ranking by relevance-score (BM25, cosine similarity, learned-to-rank models) is objective and reproducible but may not match human judgments. Incorporating user feedback (clickthrough, dwell time) improves ranking but introduces position bias (users click more on top results) and feedback loops^[7]. The system must balance algorithmic objectivity with human-grounded quality signals.

Structural–Framed Character

Search and Retrieval sits at the structural end of the structural–framed spectrum: it is a pure relational pattern, the same in any domain where it appears, and nothing about its meaning depends on a particular field's vocabulary or assumptions.

At root it is a query, a search space to be traversed, a relevance criterion that decides what matches, and a trade-off between exhaustiveness and cost—a configuration that can be defined entirely in formal terms with no reference to any human institution or norm. The same structure appears whether a database engine resolves an index lookup, a forager scans terrain for food, or a memory system retrieves a stored item, and it carries no built-in evaluative weight: a search either finds the matching items efficiently or it does not. Encountering it is a matter of recognizing a navigate-a-space-against-a-criterion pattern already present, not of importing an outside perspective. On every diagnostic, it reads structural.

Substrate Independence

Search and Retrieval is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its structure — specify a query, traverse a search space, match on relevance, and rank — is substrate-agnostic and recurs across computer science, library science, cognitive science (memory retrieval), and biology (foraging). The examples genuinely span formal algorithms like vector-space search, applied systems like web search, and biological foraging behavior, each instantiating the same structural logic rather than a borrowed metaphor. That solid, multi-substrate transfer places it firmly in the upper band at 4.

• Composite substrate independence — 4 / 5
• Domain breadth — 4 / 5
• Structural abstraction — 4 / 5
• Transfer evidence — 4 / 5

Relationships to Other Abstractions

Current abstraction Search and Retrieval Prime

Parents (2) — more general patterns this builds on

• Search and Retrieval presupposes Problem Space Prime

  Search and retrieval presupposes a problem space because locating items requires a representation specifying states, operators, and goal criteria.

• Search and Retrieval presupposes Trade-offs Prime

  Search and retrieval presupposes trade-offs because every retrieval system must balance precision, recall, and latency against each other.

Children (15) — more specific cases that build on this

• Invention (Rhetorical Canon) Domain-specific is a kind of Search and Retrieval

  Rhetorical invention is search and retrieval specialized to canvassing a structured argument space before selecting what to advance.

• Memory (Rhetorical Canon) Domain-specific is a kind of Search and Retrieval

  Rhetorical memory is search and retrieval specialized to a pre-acquired performance repertoire indexed for accurate access under live pressure.

• Search Algorithm Domain-specific is a kind of Search and Retrieval

  A search algorithm is search and retrieval specialized to machine- represented state spaces, generated successors, and frontier ordering.

▸ Show 12 more

Hierarchy paths (4) — routes to 3 parentless roots

• Search and Retrieval → Problem Space → Representation → Abstraction

Show alternative paths (3)

Neighborhood in Abstraction Space

Search and Retrieval sits in a sparse region of abstraction space (94^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely rather than landing on a neighbor.

Family — Data Integrity & Provenance Infrastructure (6 primes)

Nearest neighbors

• Spatial Indexing — 0.72
• Analogy — 0.69
• Encoding Specificity — 0.68
• Exemplar Retrieval — 0.67
• Record Reconciliation — 0.66

Computed from structural-signature embeddings · 2026-07-26

Not to Be Confused With

Search and Retrieval must be distinguished from Maintenance, though both involve system operations. Search and retrieval is the locating and accessing of existing information or resources matching a query criterion; it is transactional and episodic—given a query, find the matching item(s) and return them. Maintenance is the preserving of a system's intended operational function against degradation, wear, failure, or drift. A car's search-and-retrieval system is the diagnostic computer that locates a specific fault code; its maintenance system includes regular oil changes, tire rotations, and brake inspections that prevent faults from accumulating. A database's search-and-retrieval operation is the query engine retrieving rows matching a WHERE clause; maintenance is the backup-and-recovery system that preserves consistency after crashes, the vacuum process that reclaims unused space, and the index rebuild process that prevents performance decay. Search retrieval answers "what do I have that matches this criterion?"; maintenance answers "is the system healthy and will it continue to function?". Search-and-retrieval systems can fail gracefully (a query returns nothing or takes a long time) without impacting system function; maintenance failures (corruption of backups, failure of error detection) can silently degrade system integrity. The two operate on different timescales: search-and-retrieval is immediate (milliseconds to seconds); maintenance is periodic (hours, days, scheduled) or reactive (triggered by anomalies). They complement each other—you search to diagnose maintenance needs, and you maintain indexes to keep search fast—but they are addressing different problems.

Search and Retrieval is also distinct from Caching, the strategy of maintaining a fast-access copy of slow-to-produce or distant data to accelerate repeated access. Caching presumes the data already exists somewhere and focuses on accelerating repeated access by keeping a local, fast copy warm. Search and retrieval presumes the system starts with no knowledge of what data exists and focuses on locating what matches a query from a potentially large or unstructured dataset. A web browser's cache stores recently-visited pages locally so repeated visits are instant; a search engine's retrieval system locates documents matching a query from billions of possibilities. A CPU cache accelerates repeated access to recently-used memory locations; a database index enables retrieval of records matching a predicate from terabytes of data. The two often interact: you search to locate an item (expensive), then cache the result to accelerate repeated access (cheap). But they solve different problems. Caching solves "we know what the user wants, how do we serve it quickly?" Search-and-retrieval solves "the user has expressed a query, what dataset items actually match it?" A caching system assumes high locality of reference (repeated requests to the same data); a search system assumes queries are heterogeneous and do not repeat. A cache that doesn't contain an item is a miss (wasteful, slower than optimal); a search that doesn't locate an item is... correct (assuming the item doesn't exist). The boundary breaks down in some systems (e.g., a search engine caches recently-computed queries to avoid recomputation), but the distinction is sharp: cache is optimization via reuse; search is problem-solving via discovery.

Search and Retrieval is also not Attention, though both involve selecting which items to process from a larger set. Attention is the cognitive or organizational resource-allocation mechanism that gates what information receives deep processing, integration, and decision-making; search and retrieval is the computational mechanism that locates items matching a query. When a radiologist scans a medical image for tumors, search-and-retrieval is the process of examining pixels and identifying candidate regions; attention is the focus that the radiologist directs at regions of interest, treating them with higher scrutiny. A conversational AI system performs search-and-retrieval to locate relevant knowledge from a database; attention mechanisms in transformers gate which parts of input receive deep processing in computing responses. Organizational attention is the executive focus on strategic priorities; organizational search-and-retrieval is the business-intelligence system that locates transactions, documents, or metrics matching specified criteria. Search produces a ranked or filtered list of candidates; attention selects a subset for further processing. A search engine returns the top 10 results, but a user's attention focuses on the first 2 or 3. Search is about quantity and inclusiveness (the more relevance signal we have, the better); attention is about scarcity (time, cognitive resources, processing capacity are limited, so we focus). Both are essential: search-and-retrieval without attention would overwhelm decision-makers with too many results; attention without search-and-retrieval would require examining all items to apply focus, which is inefficient. They are compositional—search produces candidates, attention selects which to process deeply—rather than identical.

Solution Archetypes

Solution archetypes in the catalog that build on this prime — directly (this prime is a source ingredient) or as a related prime.

Built directly on this prime (18)

• Archetype Pattern Indexing: Index recurring patterns by structural signature so they can be recognized, compared, and reused across contexts.
  ▸ Mechanisms (9)
• Bounded Search Pruning: Eliminate branches of a search space only when bounds prove they cannot beat current alternatives or satisfy required thresholds.
  ▸ Mechanisms (9)
• Coarse-to-Fine Search: Search broadly at a coarse level first, then refine only the most promising regions in more detail.
  ▸ Mechanisms (8)
• Constraint-Guided Backtracking: Solve a constrained, path-dependent problem by extending a partial solution, testing it early, and undoing the latest failed commitment while preserving still-valid prior work.
  ▸ Mechanisms (7)
• Encoding–Retrieval Context Alignment: Design encoding, practice, cues, and fallback so the features available at use can recover what was learned.
  ▸ Mechanisms (16)
• Index-Based Retrieval: Create an index or retrieval structure so relevant information can be found without scanning the whole space.
  ▸ Mechanisms (12)
• Knowledge Map Navigation: Create and use a map of a knowledge domain so people can locate concepts, gaps, dependencies, and learning paths.
  ▸ Mechanisms (11)
• Landscape-Aware Search Strategy Design: Map the shape of the value surface before choosing how to search it, so effort matches the terrain instead of getting trapped by it.
  ▸ Mechanisms (9)
• Memory Palace Retrieval Indexing: Use a familiar spatial or ordered cue path as an index for reliable sequenced recall.
  ▸ Mechanisms (8)
• Nearest-Exemplar Response Reuse: Use the closest remembered or stored case as the model for the present response, while making similarity, adaptation, confidence, and exception boundaries explicit.
  ▸ Mechanisms (8)
• Operation-Weighted Data Structure Design: Choose the information structure around the real operation mix, making lookup, update, traversal, storage, consistency, and maintenance tradeoffs explicit instead of accidental.
  ▸ Mechanisms (11)
• Predictive-Cue Wayfinding Design: Make local cues honestly predict what lies down each path so agents can choose, continue, or recover without needing a complete map.
  ▸ Mechanisms (9)
• Problem Space Mapping: Map the states, actions, constraints, and goals of a problem so exploration becomes deliberate rather than ad hoc.
  ▸ Mechanisms (9)
• Progressive Narrowing: Narrow a broad option space step by step until a stable choice, design, diagnosis, explanation, or bounded issue set remains.
  ▸ Mechanisms (10)
• Registry-Mediated Discovery: Put a maintained discovery registry between agents and changing counterparts so stable names resolve to current locations, interfaces, or contact records instead of hard-coded references.
  ▸ Mechanisms (10)
• Search Space Pruning: Reduce an overwhelming search space by eliminating candidates or regions that cannot plausibly satisfy constraints or improve the outcome.
  ▸ Mechanisms (12)
• Solution Space Bounding: Bound a potentially unbounded or enormous solution space so search becomes possible.
  ▸ Mechanisms (9)
• Strategic Caching: Store high-value reusable results near where they are needed so repeated retrieval or computation becomes faster and less costly.
  ▸ Mechanisms (8)

Also a related prime in 38 archetypes

• Access-Optimized Redundant Representation: Create a governed redundant representation around a proven access path, keep one authority and an explicit derivation, bound divergence, verify the benefit, and make refresh, repair, schema change, privacy, and retirement part of the design.
• Accumulation Compaction: Compress accumulated layers or records so history remains usable without overwhelming present operation.
• Activation Decay Measurement: Treat priming as a fading state: measure its useful lifetime, set an action or refresh window, and stop relying on it after it expires.
• Adaptive Mutation Rate Management: Treat deliberately introduced variation as a tunable control variable: increase it when the system needs exploration and reduce it when the system needs stability, safety, or convergence.
• Advantageous Repositioning: Gain advantage by moving to a better position in the option, terrain, timing, information, or institutional space instead of fighting the same contest from a worse position.
• Cascaded Hierarchical Recognition: Recognize complex cases by moving attention through a hierarchy of coarse filters and fine discriminators instead of trying to inspect every possible feature at once.
• Chunked Information Design: Group information into meaningful chunks so it can be understood, remembered, retrieved, and acted on more easily.
• Constraint Formulation: Turn implicit limits, requirements, and prohibitions into explicit constraints that shape the feasible solution space.
• Constraint Propagation and Decoupling: When constraints bind a problem into an unwieldy whole, propagate their implications first, then solve only the reduced and justified subproblems that remain.
• Cross-Axis Product Space Design: Define independent axes, list each axis's allowed choices, form the cross-product, and govern which cells are valid, covered, sampled, or deliberately excluded.

▸ Show 28 more

Notes

Search and Retrieval is held at High confidence. Foundational abstraction in computer science (databases, search engines, operating systems), cognitive science (memory retrieval), and biology (foraging). The vector space model and inverted indexes are canonical formalisms with decades of empirical validation. Modern instantiations (Elasticsearch, vector databases, learned indexes, semantic embeddings) build on classical information retrieval, demonstrating lasting architectural relevance.

References

[1] Salton, G., & McGill, M. J. (1983). Introduction to Modern Information Retrieval. McGraw-Hill. ↩

[2] Kraska, T., Beutel, A., Chi, E. H., Dean, J., & Polyzotis, N. (2018). "The case for learned index structures." Proceedings of the 2018 International Conference on Management of Data (SIGMOD). ↩

[3] Robertson, S., Walker, S., Jones, S., Hancock-Beaulieu, M. M., & Gatford, M. (1995). "Okapi at TREC-3." Proceedings of the Third Text REtrieval Conference (TREC-3). ↩

[4] Brin, S., & Page, L. (1998). "The anatomy of a large-scale hypertextual web search engine." Computer Networks and ISDN Systems, 30(1–7), 107–117. ↩

[5] NIST TREC. Text REtrieval Conference. https://trec.nist.gov/. ↩

[6] Dean, J., & Ghemawat, S. (2004). "MapReduce: Simplified data processing on large clusters." OSDI 2004[^distributed-search]. ↩

[7] Joachims, T., Grover, A., & Ping, B. (2017). "Deep learning with differential privacy." Journal of Machine Learning Research, 52, 310–328. ↩

[8] Schroff, F., Kalenichenko, D., & Philbin, J. (2015). "FaceNet: A unified embedding for face recognition and clustering." IEEE Conference on Computer Vision and Pattern Recognition (CVPR).