Explain semantic and exact-match retrieval

Retrieval selects candidate items from a collection for a query. The collection might contain help articles, code snippets, product descriptions, or document passages. A result usually contains an item identifier, a ranking score, and enough metadata to locate or display the item.

Suppose a learning platform receives the query “how to stop duplicate requests in React”. Its search might retrieve a lesson about request deduplication even when the lesson title uses different words. The retrieved lesson is a candidate source. Whether it actually solves the learner's problem still needs assessment.

Retrieval is also part of RAG, where selected material is supplied to a generator. This topic focuses on finding and ranking candidates. The RAG & Data Pipelines topic covers ingestion, grounding, citations, and the document lifecycle in greater depth.

An embedding model converts an input into a vector of numbers. A text embedding model is trained so that useful relationships between texts can be reflected in vector comparisons. The exact relationship depends on the training objective and model.

A vector can have hundreds or thousands of dimensions. Individual coordinates usually have no simple human-readable meaning. An embedding is not a list of keywords, and it is not a lossless encoding from which the original document can reliably be reconstructed.

Text with similar meaning may receive similar vectors, but the result can still confuse negation, identifiers, dates, or specialized vocabulary. Test the intended domain rather than assuming that semantic similarity guarantees relevance.

Stored document vectors and incoming query vectors must come from compatible representations. Equal dimensionality alone does not establish compatibility. Different models can assign completely different meanings to their coordinates.

Some retrieval models expect different instructions or encoders for queries and documents. Follow that model's supported query and document workflow. Store the embedding configuration with the index so a later query cannot accidentally use a different representation.

When the embedding model changes, document vectors generally need to be regenerated into a compatible index. Changing only the query encoder can make rankings meaningless even when the search engine accepts the vector length.

Three common comparisons are dot product, cosine similarity, and Euclidean distance. Dot product multiplies matching coordinates and sums them. Cosine divides the dot product by the two vector lengths, emphasizing direction. Euclidean distance measures straight-line separation.

Higher similarity usually ranks first, while lower distance usually ranks first. Verify the score convention of the actual engine. Some APIs return transformed distances, so the displayed number may not be the mathematical quantity named in a textbook.

For nonzero unit-length vectors, dot product equals cosine similarity. Normalization is therefore a meaningful configuration choice. Use the metric expected by the representation model, and apply normalization consistently.

BM25 is a common lexical ranking method. It considers factors including term frequency, term rarity, and document length. A basic term-overlap script is a useful teaching baseline, but it does not reproduce BM25.

Hybrid retrieval combines results from different methods. A useful candidate might appear in one method's results even when the other misses it. Combining methods still requires evaluation because additional candidates can introduce irrelevant material.

An index is a data structure used to search a collection efficiently. A vector index searches stored vectors. A full-text index supports lexical retrieval. A database may provide both capabilities alongside metadata storage.

Top-k means returning up to k ranked results. It does not mean that k relevant results exist. If the collection has no useful answer, nearest-neighbor search can still return the closest available vectors.

Exact vector search compares the query against every eligible vector according to the selected metric. Approximate nearest-neighbor (ANN) search trades some exactness for efficiency. Its effect depends on the index, configuration, data, and query distribution.

Metadata can restrict results to a language, tenant, document type, or current publication state. For example, a learning platform may limit a query to published React lessons that the current learner can access.

Authorization constraints must be applied by trusted application logic. A query string should not grant access to private records. Filtering only after a very small candidate set is selected can also leave fewer than k results, even when additional eligible items exist elsewhere in the collection.

A search score expresses a ranking signal. It does not establish permission to display an item.

A long document can be represented as smaller passages. Each passage should retain its parent identifier and useful context. A passage about cancellation may be useful even when the entire document covers many unrelated API features.

Passage boundaries affect retrieval. A fragment that omits the name of the function or the condition under which advice applies can be difficult to retrieve and interpret. Detailed chunking strategies belong in RAG & Data Pipelines, but candidates should recognize their effect on search quality.