Retrieval Pipeline — AI Glossary

Reference details

Related terms

• Indexing Pipeline
• Advanced RAG
• Modular RAG

How it works

1. The query is normalised and often rewritten or expanded, since the text a user typed and the text that retrieves well are frequently not the same string.
2. One or more retrievers produce a candidate set — dense vector search, sparse lexical search, or both with their scores fused into a single ranking.
3. Metadata filters are applied to restrict candidates by permission, recency or document type, and whether this happens before or after the vector search changes both recall and latency.
4. A cross-encoder reranker then scores each surviving candidate against the query jointly, which is more accurate than vector similarity and too slow to run over the whole corpus.
5. Survivors are compressed or deduplicated to fit the context budget, which is the last stage at which content is discarded.
6. Every stage only ever narrows the set, so recall is non-increasing from one end to the other: a document dropped at filtering cannot be recovered by reranking, which is why the stages must be measured in order rather than from the output backwards.

Example: Where quality is actually lost

A pipeline is query rewriting → search → filter → rerank → compress → generate. When answers are poor, the instinct is to change the model, but the loss is usually upstream: chunks split mid-sentence, a metadata filter excluding valid documents, or reranking absent so the model reads whatever the index ranked first. Measure each stage before touching the last one.

Commonly confused with

The pipeline is not the vector database. The database is one stage — storage and nearest-neighbour search. Chunking, query processing, filtering, reranking and compression sit around it, and most retrieval quality problems live in those stages rather than in the index.

When to use it

Reach for it when:

• Retrieval quality is inconsistent and you need to isolate which stage is failing
• You are combining several retrieval strategies and need one order
• Different document types need different chunking or filtering

Reach for something else when:

• The corpus is small and simple — stages add latency and failure modes
• You are adding stages without measuring each one; complexity hides the real fault

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Retrieval Pipeline — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path