[Design] Search Embedded Aggregation

[MrPresent-Han]

Summary

Propose a design for Search Embedded Aggregation: extend Milvus vector search to return per-group top-K hits and per-group aggregation metrics (avg / sum / count / max / min) in a single request, aligning with Elasticsearch's terms + top_hits + metric sub-aggregation combo.

Motivation

Milvus today supports single-field group_by_field + group_size for grouping search results, but the most common vector-search analytics requests — e.g. "for each document, return the top-2 most similar chunks AND the average / max similarity / chunk count of that document" — cannot be expressed in a single search call. Users must issue two separate requests (search + query aggregation) and stitch results on the client, losing both ergonomics and latency.

Gap Analysis

#	Capability	ES	Milvus master	Gap
R1	Single-field grouping	yes	yes	covered
R2	Per-group top-K hits	yes (top_hits.size)	yes (group_size)	covered
R3	Multi-field flat grouping (ES multi_terms)	yes	segcore foundation only (commit 247d55bb36)	Proxy / QueryNode / SDK
R4	Per-group metrics (avg/sum/count/max/min/...)	yes (20+)	no	core gap
R5	Non-vector sort inside a group	yes	no	uncovered
R6	Hierarchical (nested) grouping	yes	no	uncovered

R4 is the dominant gap.

Scope

• R3 — Multi-field flat grouping aligned with ES multi_terms (top segcore work already landed in commit 247d55bb36, Go layer still missing).
• R4 — Leaf-level per-group metrics (count / sum / avg / max / min in Phase 1; cardinality / percentiles later).
• R5 — Non-vector group-internal sort.
• R6 — 2-level nested grouping via proxy-side flattening over composite keys, with group_count_safe_factor and metric_safe_factor knobs controlling two orthogonal approximation errors. Leaf-only metrics in Phase 1.

Design Document

A full MEP has been drafted: design_docs/20260413-search_embedded_agg.md (PR incoming).

Related Work

• Commit 247d55bb36 — feat: support search embeded group by — segcore composite-key foundation already merged locally on branch support-embeded-groupby-master.

cc @MrPresent-Han