Hybrid Retrieval

Hybrid retrieval combines lexical retrieval and embedding retrieval. It is useful when exact terms, symbols, acronyms, paper titles, identifiers, or code names matter, but semantic paraphrase also matters.

The combined score can be written as:

$$s(q,d)=\lambda s_{\mathrm{lex}}(q,d)+(1-\lambda )s_{\mathrm{emb}}(q,d)$$

where $s_{\mathrm{lex}}$ is a lexical score, $s_{\mathrm{emb}}$ is an embedding similarity score, and $\lambda \in [0,1]$ controls the mixture.

Because the two scores often have different scales, practical hybrid retrieval usually normalizes or ranks before combining:

$$s^{\prime }(q,d)=\lambda \mathrm{norm}(s_{\mathrm{lex}}(q,d))+(1-\lambda )\mathrm{norm}(s_{\mathrm{emb}}(q,d))$$

Why It Matters

• Lexical retrieval catches exact names, formulas, filenames, and rare terms.
• Embedding retrieval catches paraphrases and conceptual similarity.
• Reranking can combine both candidate sets with a more expensive model.
• Metadata filters can enforce scope before ranking.

Pipeline

$$Q\to \{\text{lexical candidates},\text{embedding candidates}\}\to \text{merge}\to \text{rerank}\to \text{context packing}$$

Candidate Merge

Hybrid retrieval is often more robust when lexical and dense retrieval produce separate candidate sets first.

$$C={\mathrm{TopK}}_{\mathrm{lex}}(q)\cup {\mathrm{TopK}}_{\mathrm{emb}}(q)$$

Then a merge or reranker orders $C$.

Merge method	Use when	Risk
score interpolation	scores are calibrated or normalized	one score family dominates
rank fusion	scales are not comparable	loses magnitude information
reciprocal rank fusion	robust simple baseline	can overvalue shallow matches
learned reranker	enough labels or strong cross-encoder exists	slower and harder to debug

Reciprocal rank fusion is a common simple form:

$$s_{\mathrm{RRF}}(d)=\sum _{r\in \mathcal{R}}\frac{1}{k_0+{\mathrm{rank}}_r(d)}$$

where $\mathcal{R}$ is the set of retrieval runs.

When Hybrid Helps

Query type	Lexical signal	Embedding signal
paper title or method name	exact phrase	related discussion
code symbol or file path	exact token	surrounding explanation
formula or acronym	literal match	expanded meaning
broad concept question	weak	semantic neighborhoods
public wiki navigation	page title and aliases	conceptual adjacency

Hybrid retrieval is useful for LLM Wiki because a user may search for Xid, RMSD, JEPA, or Slurm QOS, where exact terms and semantic explanation both matter.

Debugging Slices

Failure	Check
exact document missing	lexical candidate set too small or tokenization issue
semantically right but wrong version	metadata/date filter missing
many duplicates	chunking too fine or merge lacks diversity
top result is related but not evidential	reranker objective too broad
rare term ignored	embedding model did not preserve identifier

Checks

• Is the query term-sensitive, semantic, or both?
• Are lexical and embedding scores normalized before interpolation?
• Does the merge preserve source diversity rather than many near-duplicates?
• Is reranking evaluated separately from initial recall?
• Are metadata filters applied before retrieval when scope is known?
• Does the final context cite the exact retrieved evidence?
• Are lexical recall, embedding recall, and reranker quality evaluated separately?
• Does merge logic preserve both exact-match and semantic candidates?

Related

• Embedding retrieval
• Retrieval
• Reranking
• Ranking metrics
• Context packing
• Retrieval-augmented generation