Embeddings and rerankers

What do embeddings represent?

An embedding is a numerical representation used to compare items such as text passages and queries. A text embedding model maps each input into a vector. A search system can compare those vectors to find passages likely to be relevant even when the wording differs from the query.

The representation is useful for retrieval, but it is not a generated answer. The application still needs to decide which results to present or pass into a language model. It also needs to retain the source text and identifiers that connect a retrieved result to its original document.

Where does a reranker fit?

A reranker takes a query and candidate passages, then scores or orders those candidates by relevance. It usually operates after an initial retrieval step. That first step narrows a large collection to a manageable candidate set; reranking helps choose a more useful order within that set.

The two stages solve different problems. Embeddings support finding candidates across an indexed collection. Reranking examines the candidates selected for a particular query. A reranker cannot recover a document that the earlier stage never supplied, so evaluate retrieval coverage as well as final ordering.

IndicStack's on-request catalog includes embedding and reranking models for search workloads, including BGE and Qwen3 variants. These are deployment options for application teams; their presence in the catalog does not imply that a chat-completions request is the correct interface for a specialist model.

How do I evaluate a retrieval pipeline?

Build a small set of realistic questions with the documents that should answer them. Check whether the expected passages appear among the candidates, then inspect their positions after reranking. Include queries that have no useful answer in the collection so the application can avoid presenting weak matches as evidence.

Document preparation matters. Chunk size, headings, metadata and duplication can change what retrieval returns. Preserve source references through processing, and recheck results when the collection or embedding model changes.

Finally, evaluate the answer stage separately. Supplying a relevant passage does not guarantee that a language model will use it correctly. Review both the retrieved evidence and the generated response, and keep enough traceability to locate where a failure began.

Related terms