Skip to content

CONCEPT Cited by 1 source

Top-down vs bottoms-up generation

Definition

Top-down vs bottoms-up generation is the architectural choice axis for LLM-based content generation where the output is a structured hierarchy (page with themed sections, document with chapters, playlist with themed segments):

  • Top-down generation — first generate the structure (themes, sections, ordering), then generate the items that belong inside each structural slot.
  • Bottoms-up generation — first generate all items, then cluster or group them into a structure.

The choice is load-bearing for three properties: per-structural- slot personalization, cross-structural-slot cohesion, and adaptability to shifting upstream objectives.

Comparison

Axis Top-down Bottoms-up
Generation order Structure → items Items → structure
Cross-section cohesion Enforceable at structure step Emergent from clustering
Personalization Per-section, structured Per-item, structure-agnostic
Adaptability to objectives Re-prompt Phase 1 Re-cluster Phase 2
Retrieval-augmentation opportunity Phase 2 consumes Phase 1 context Harder — no pre-committed structure to retrieve against
Fine-tune-to-objective cost Phase-1 prompt edit Whole-task retrain
Modelling complexity Multiple focused models One broad model
Computational cost Higher per-call (cascade) Lower per-call (single model) but context-larger

Canonical wiki instance — Instacart Shopping Hub (2026-02-26)

Source: sources/2026-02-26-instacart-our-early-journey-to-transform-discovery-recommendations-with-llms

Instacart explicitly benchmarks both paradigms for their Shopping Hub rebuild and picks top-down. The post's framing:

"Bottoms-up generation: directly generate all possible products to serve to a user, then cluster and organize them into placements."

"Top-down generation: begin by generating ordered placements to structure the entire page, then generate products per placement."

Bottoms-up evaluation in the post:

  • Flexibility — less constrained recall, can surface items the structure didn't anticipate. "Deep flexibility with less constrained recall."
  • Latency + catalog turnover — difficulties in real-world serving. Generating all products at serving time is prohibitive.
  • Adaptability risk — "with a much broader modeling task, it can be difficult to ensure generated products meet a diverse set of page requirements and intents, and may require significant fine tuning efforts as needs evolve."

Top-down evaluation in the post:

  • Balances personalization, cohesion, and adaptability. "To best balance personalization, cohesion, and adaptability, we landed on a top-down, cascaded approach."
  • Cohesion-first. Themes are first generated + ordered against user relevance, cross-theme cohesion, and other business goals.
  • Cascade-amenable. Phase 1 emits structural decisions + derived signals → Phase 2 consumes them + retrieves candidates → Phase 3 filters. The structure is what makes the cascade possible.

Instacart's canonical example — building two placements: Breakfast Staples + Health-Conscious Snacks. Bottoms-up generates a raw sequence of products then clusters them. Top-down generates "Easy Pasta Night" + "Gourmet Salad Fixings" first, then generates products per theme.

Why top-down wins for hierarchical content surfaces

Three structural reasons implicit in the post:

  1. Structure-first locks in personalization. Phase 1 sees the user context + all business objectives and chooses the structure accordingly. Bottoms-up would have to embed this in either the item generator (hard) or the clusterer (limited by how items cluster, not by what themes the user wants).
  2. Structure-first opens the door to RAG + other cost-shape primitives. Phase 1's freeform concept emission lets Phase 2 do RAG candidate pruning. Bottoms-up has no equivalent intermediate signal to retrieve against.
  3. Structure-first is cheaper to re-prompt for new business objectives. Swap the Phase-1 prompt for seasonal pushes, novel vs relevant balance, etc. Bottoms-up requires retraining the item generator's objective.

When bottoms-up wins

  • Structure is emergent rather than designed. Music discovery from implicit user listening patterns, where thematic coherence emerges from the catalog rather than being imposed.
  • Recall is the dominant concern. If the cost of missing a great item is worse than the cost of a slightly-chaotic page, generate items first.
  • Users don't expect structured sections. An infinite-scroll surface doesn't need a themed structure.
  • Item catalog is small. Generating all items then clustering is tractable only at modest catalog size.

Relation to other design axes

Seen in

Last updated · 517 distilled / 1,221 read