A multi-stage retrieval pipeline is the query language for unstructured data. Like SQL composes WHERE, ORDER BY, LIMIT, and JOIN for structured data, multi-stage retrieval composes filter, sort, reduce, enrich, and apply stages for multimodal content. Each stage takes the previous stage's output as input, progressively narrowing and enriching the result set.

How It Works

A pipeline is defined as an ordered list of stages. The first stage (usually a filter) searches across an embedding space to produce an initial candidate set. Subsequent stages narrow, reorder, sample, or enrich those candidates. For example: face search (847 candidates) → logo filter (23) → sentiment sort (23 reordered) → top-k reduce (5) → brand context enrich (5 enriched results).

The SQL Analogy

Filter stages are like WHERE clauses: they narrow the candidate set by embedding similarity or metadata
Sort stages are like ORDER BY: they re-rank results by weighted combinations of scores
Reduce stages are like LIMIT and DISTINCT: they control output size via sampling or deduplication
Enrich stages are like JOIN: they attach data from other collections (the semantic join)
Apply stages are like INSERT INTO...SELECT: they write results or trigger side effects

Best Practices

Start with broad filter stages and narrow progressively; don't over-constrain early
Use sort stages to combine multiple ranking signals with explicit weights
Always include a reduce stage to control output size
Use enrich stages for cross-collection context instead of client-side merging
Test pipelines with different stage orders, since the sequence affects both quality and latency

Retriever documentation: /docs/retrieval/retrievers
Retrieval Cookbook: /docs/retrieval/cookbook
Blog: Multi-Stage Retrieval Pipelines - /blog/multi-stage-retrieval-pipelines

Managed Mixpeek

Put multimodal search to work

Connect a bucket and Mixpeek runs the whole multimodal search pipeline for you: extraction, indexing, and search over your own objects. No models to wire up, nothing to host.

Start with Managed

MVS · bring your own