Retrieval, Hybrid Search, and Reranking

Key jargon

Term	Plain-language meaning
Sparse retrieval	Term-based search that rewards lexical overlap, commonly using BM25.
Dense retrieval	Vector search that compares learned embeddings.
Hybrid search	Combining sparse and dense retrieval signals.
Reranking	Applying a more expensive relevance scorer to a smaller candidate set.

Key concepts

Sparse search is strong for exact names and identifiers; dense search helps with paraphrases and semantic similarity.
A reranker improves ordering only among candidates already retrieved, so candidate recall still matters.

Concept map

flowchart LR
    A["Run sparse and dense search"] --> B["Fuse candidate rankings"]
    B --> C["Rerank shortlist"]
    C --> D["Return evidence with scores"]

Method	Strength	Weakness
Lexical/BM25	Exact names, identifiers, rare terms	Misses paraphrases
Dense vector	Semantic similarity	Can blur exact constraints
Metadata filter	Tenant, date, type, policy	Depends on correct metadata
Hybrid	Balances exact and semantic signals	Requires score fusion/tuning
Reranker	Better ordering of candidates	Adds latency and model risk

Retrieve broadly enough to include the answer, then rerank/filter narrowly enough for useful context. Tune on real questions, including “no answer” cases.

Exercise

Create queries containing exact error codes, synonyms, dates, and ambiguous terms. Compare lexical and vector results, then define a simple hybrid rule.

Metrics

Recall@k, precision@k, mean reciprocal rank, nDCG, policy-filter correctness, freshness, latency, and cost. Metrics require labeled relevance judgments.