Knowledge Graph Index

The triplestore is the specialized database engine designed to store, manage, and query RDF triples (subject-predicate-object). It is optimized for graph traversal operations rather than table joins.

• Core Function: Persists the graph data and provides a SPARQL endpoint for querying. SPARQL is a SQL-like query language for graphs.
• Indexing Strategies: To enable fast lookups for complex graph patterns, triplestores create multiple indices over permutations of the triple (e.g., SPO, SOP, PSO, OSP). This allows efficient answering of queries like "find all subjects with predicate X and object Y."
• Examples: Open-source triplestores include Apache Jena Fuseki and RDF4J, while commercial graph databases like Neo4j (using Cypher query language) also serve this role.

A specialized database optimized for storing high-dimensional vector embeddings and performing approximate nearest neighbor (ANN) search. It is the primary engine for semantic similarity search in RAG systems.

• Core Function: Enables queries like "find documents most similar to this embedding."
• Contrast with KG Index: While a vector store excels at finding semantically similar items, a knowledge graph index excels at traversing explicit relationships (e.g., multi-hop queries).
• Common Use: Often used in tandem with a KG index in a hybrid retrieval architecture.

The foundational data structure for keyword-based (lexical) search. It maps each unique term in a corpus to a list of documents where that term appears.

• Mechanism: Enables fast Boolean and term-frequency based queries (e.g., "documents containing 'transformer' AND 'attention'").
• Role in Hybrid Search: Provides the sparse, exact-match component that complements the semantic understanding of vector search. BM25 is the classic ranking algorithm used with inverted indices.
• Limitation: Cannot handle synonyms or conceptual queries without the term being literally present.

A retrieval strategy that combines the results of sparse (keyword) and dense (vector) search methods to improve overall recall and precision.

• Implementation: Typically involves querying both an inverted index (using BM25) and a vector store, then fusing the ranked results using a weighted scoring function like Reciprocal Rank Fusion (RRF).
• Advantage: Mitigates the weaknesses of each approach: lexical search's vocabulary mismatch and vector search's potential for semantic drift.
• Advanced Pattern: Can be extended to a "three-tower" hybrid incorporating a knowledge graph index for relationship-based retrieval.

The natural language processing task of identifying and classifying named entities (e.g., persons, organizations, locations) within text. It is a critical preprocessing step for populating a knowledge graph.

• Function: Transforms unstructured text into structured entity mentions, which become candidate nodes for the graph.
• Relation Extraction: Often paired with relation extraction to identify how entities are connected, forming the edges of the graph.
• Tooling: Implemented using models like spaCy's NER, or foundation model APIs.

A database management system designed specifically to store, query, and manipulate graph structures (nodes, edges, properties). It provides the transactional and persistence layer for a knowledge graph.

• Query Language: Uses declarative languages like Cypher (Neo4j) or Gremlin (Apache TinkerPop) for expressive graph traversals.
• Contrast with Index: The knowledge graph index is often an optimized, query-focused layer built on top of or alongside a graph database to accelerate specific retrieval patterns for AI agents.
• Examples: Neo4j, Amazon Neptune, JanusGraph.

A state-of-the-art, graph-based algorithm for Approximate Nearest Neighbor (ANN) search in high-dimensional spaces. It is the most common indexing method used in modern vector stores.

• Mechanism: Constructs a hierarchical graph where long-range connections enable fast, log-time search complexity.
• Relevance: While HNSW indexes vector embeddings for similarity, a knowledge graph index structures discrete entities and relations for logic. They solve different but complementary retrieval problems.
• Ubiquity: The default index in libraries like FAISS and databases like Weaviate and Qdrant.