Recuperação densa versus esparsa: compromissos

Modern retrieval systems rely on two distinct signals: dense retrieval encodes meaning into continuous vector spaces, while sparse retrieval counts exact term occurrences. These signals are complementary, not interchangeable. Understanding their individual strengths and weaknesses is the first step toward building a system that uses both effectively.

Dense retrieval maps both the query and each document into a high-dimensional vector using a neural encoder. Similarity is measured by cosine distance or dot product between vectors. Because the encoder was trained on large text corpora, semantically related phrases end up near each other in vector space even if they share no common words — this is the key advantage of dense retrieval.

Dense models struggle with rare terms that were underrepresented during encoder training. A query containing a specific product model number like RTX-4090-Ti-OC, a medical drug name, or a proprietary internal identifier will often fail to match the correct document because the encoder has no learned representation for that token sequence. The vector just lands somewhere unhelpful in embedding space.

BM25 (Best Matching 25) is a probabilistic ranking function that scores documents based on how often query terms appear in the document, normalized for document length and dampened by term frequency saturation. It produces a sparse score vector — most dimensions are zero because documents contain only a small fraction of the vocabulary.

BM25 excels at queries involving exact technical terms, product names, error codes, and numeric identifiers that should match precisely. A query for ORA-01017 (an Oracle error code) will rank documents containing that exact string far above documents that merely discuss database authentication in general terms. This is impossible for a dense model that has never seen that specific code.

BM25 is blind to semantic paraphrasing. A document about 'automobile engine repair' will score zero for a query about 'car motor maintenance' because none of the exact words overlap. This vocabulary mismatch problem, sometimes called the lexical gap, means pure keyword search misses huge amounts of relevant content that simply uses different words to express the same idea.

Benchmarks on BEIR, MS MARCO, and enterprise Q&A datasets consistently show that hybrid retrieval outperforms either dense or sparse alone by 5-15 percent on NDCG@10. The improvement is largest on datasets with a mix of factual lookups (where BM25 helps) and paraphrase queries (where dense embeddings help). No single retrieval method dominates across all query types.

You can predict which retriever will perform better by analyzing the query type. Dense retrieval wins on conceptual questions, paraphrases, and broad topic queries. BM25 wins on queries containing proper nouns, version numbers, code snippets, acronyms, and rare technical terms. Hybrid always wins when query type is unknown in advance — which is almost always true in production.

Combining dense and sparse results is non-trivial because their scores are on incompatible scales. Cosine similarity produces values between -1 and 1, while BM25 produces unbounded positive scores that depend on corpus size. You cannot simply add them. The standard solution is to use rank-based fusion rather than score-based fusion — merging ranked lists instead of raw scores.

Use dense-only retrieval when your corpus is in a narrow domain with consistent vocabulary and you need semantic generalization across paraphrases. Use BM25-only when queries are primarily lookup-style with exact identifiers and your dataset is small enough that brute-force is feasible. Use hybrid in all production RAG systems where query types vary — the overhead is modest and the recall improvement is significant.

Dense retrieval requires GPU-accelerated approximate nearest neighbor search or a vector database, which adds infrastructure cost. BM25 runs entirely on CPU with an inverted index and is extremely fast. Hybrid retrieval requires both infrastructure components plus a fusion step. The added complexity is justified by the recall improvement for most production use cases, but must be weighed against your infrastructure budget.

Test your understanding of dense versus sparse retrieval trade-offs from this lesson.

In this lesson you learned: dense retrieval captures semantic meaning but fails on rare exact terms, BM25 sparse retrieval handles exact keywords but misses paraphrases, and hybrid retrieval consistently outperforms either method alone across diverse query types. Their scores are incompatible and must be merged via rank fusion rather than score addition. Next up we implement BM25 keyword search in Python.