Retrieval Pipeline

What is a retrieval pipeline?

A retrieval pipeline is the sequence of stages that turns an information need into selected evidence. It may prepare the query, search one or more indexes, apply filters, combine candidates, rerank results, and assemble context for a downstream model or application.

Retrieval is therefore not always one search call; it can be a controlled multi-stage process.

Common pipeline stages

A practical pipeline may include:

1. Interpret or rewrite the query.
2. Apply tenant, permission, date, or product filters.
3. Run sparse, dense, or hybrid retrieval.
4. Merge and deduplicate candidates.
5. Rerank by relevance and other constraints.
6. Select a diverse evidence set.
7. Attach source metadata and build model context.

For example, a policy assistant may rewrite a conversational request, filter to the correct region and effective date, run hybrid search, rerank passages, and return the top evidence with document links.

Retrieval pipeline versus RAG

A retrieval pipeline finds and prepares evidence. [RAG](/glossary/retrieval-augmented-generation/) combines retrieved evidence with model generation.

The same retrieval pipeline can support search results, recommendations, memory recall, or analytics without generating an answer. RAG uses the pipeline's output as model context.

Pipeline versus individual techniques

Query rewriting, dense retrieval, sparse retrieval, metadata filtering, and reranking are stages or techniques within a pipeline. They solve different problems and should be evaluated separately.

If the correct document never enters the candidate set, reranking cannot recover it. If retrieval is strong but the model ignores evidence, the generation stage—not the retrieval pipeline—may be failing.

Quality and observability

Builders should record the original query, rewrite, filters, candidate sources, scores, reranking changes, and final selected chunks. This helps diagnose where relevant evidence was lost.

Evaluation should measure candidate recall, ranking quality, freshness, access correctness, latency, and context usefulness. A final-answer score alone hides which stage needs improvement.

Why it matters

Retrieval pipelines make evidence selection deliberate and testable. They help agents use limited context efficiently while preserving permissions and provenance.

Complexity should match the problem. A simple collection may need one search step; a high-stakes enterprise system may require several stages and explicit fallbacks. See [Memory in AI Agents](/memory-in-ai-agents/) for how retrieval supports external memory.