LLM-Powered Content Embeddings for Multi-Vertical Search and Recommendations

doordash · E-commerce · 2026 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Multi-vertical search and recommendations

DoorDash needed better semantic retrieval and personalization across restaurants, grocery, retail, and gifting, where sparse metadata and engagement-only embeddings missed catalog meaning and cold-start intent.

Constraints:

• Behavioral signals are proxies for meaning and are sparse in transactional, short-session commerce.
• Catalogs span multiple verticals with different metadata quality and engagement density.
• Embeddings had to support both large offline catalog backfills and low-latency online query embedding.
• The system had to improve search, recommendation, and cold-start discovery without relying only on accumulated clicks.

Architecture

DoorDash built a content-first embedding pipeline that turns LLM-generated merchant and item profiles into reusable dense representations for search and recommendation surfaces. Image and metadata inputs become comprehensive text profiles, Metaflow incrementally refreshes changed catalog entities, gemini-embedding-001 with 256-dimensional MRL encodes entities and queries, and downstream search/recommendation systems use those embeddings for semantic retrieval, item-level RAG/reranking, and personalized homepage carousels.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  LLM_profiles["LLM-generated profiles"] --> Embedding_pipeline["Embedding inference pipeline"]
  Item_images["Item image descriptions"] --> LLM_profiles
  Embedding_evaluation["Embedding model evaluation"] --> Gemini_model["Gemini embedding model"]
  Gemini_model --> Embedding_pipeline
  Embedding_pipeline --> Embedding_index["Embedding index"]
  Gemini_model --> Online_query["Online query embedding"]
  Online_query --> Semantic_search["Semantic search retrieval"]
  Embedding_index --> Semantic_search
  Semantic_search --> Qwen_reranker["Qwen reranker"]
  Embedding_index --> Homepage_retrieval["Homepage carousel retrieval"]
  Homepage_retrieval --> Store_ranker["Existing store ranker"]
  Qwen_reranker --> Search_presentation["Search result presentation"]
  Store_ranker --> Search_presentation

Components

Data flow

Merchant/item metadata and, when present, images are transformed into natural-language profiles. The embedding pipeline refreshes only entities whose underlying content changed, writes embeddings into persistent storage/indexes, and serves them to entity similarity, embedding-based retrieval, item-level search RAG/reranking, and homepage generative carousel retrieval.

Why this architecture

DoorDash chose content-first embeddings because commerce sessions are intentful and sparse, making engagement-only representations weak for tail queries and cross-vertical coverage. LLM profiles improve input representation, then embeddings, rankers, and sequence models bend the semantic space toward DoorDash's transactional objectives.

Models used: gemini-embedding-001 · text-embedding-005 · MiniLM · Qwen embedding models · Qwen 3 Rerank · vision-language model · LLMs

Scale

• The system spans DoorDash food, grocery, retail, and gifting verticals.
  “This problem spans every DoorDash vertical — including food, groceries, retail, and gifting, with each holding catalog richness that sparse metadata flattens away.”article, stated
• DoorDash adopted 256-dimensional MRL embeddings to balance quality and index efficiency.
  “we adopted gemini-embedding-001 with 256-dimensional output -- leveraging MRL — using SEMANTIC_SIMILARITY task type for entity-entity comparisons”article, stated
• Semantic search reduced null search rate by 3.65% and raised core search session conversion by 0.66%.
  “The null search rate reduction is particularly telling — 3.65% fewer searches return nothing useful, which is precisely the tail-query scenario”article, stated

  “Core search session CVR is increased by +0.66%”article, stated
• Homepage generative carousels increased order rate by 2.4% and improved offline precision@10 from 68% to 85%.
  “Consumer homepage order rate increased by 2.4% relatively; consumer reorder rate in the previous seven days increased by+0.164% relatively”article, stated

  “Offline precision@10 on the homepage improved 68% to 85%.”article, stated

Alternatives considered (roads not taken)

Behavior-only embeddings from co-visitation and engagement

Rejected because: Clicks and co-visitation were only proxies and did not capture identity, context, intent, or semantic distinctions in DoorDash's catalog.

Relying on metadata or human labels without LLM profile generation

Rejected because: Content encoders were historically constrained by metadata richness and label quality; LLM profiles changed the data-quality bottleneck.

Swapping to a better embedding model without improving input representation

Rejected because: DoorDash's offline evaluation showed data quality moved performance more than model choice alone.

Single consumer text-profile embedding for personalization

Rejected because: Consumer intent varies by time, location, occasion, and trajectory, so a static text vector averaged away useful distinctions.

Patterns

fine-tuning-adaptation (modeling)

DoorDash used domain-specific LLM-generated text profiles as the main adaptation layer before embedding, improving representation quality for merchants and items without custom-training every retriever.

hybrid-search (retrieval)

The search stack combines existing lexical/store retrieval with embedding-based retrieval and item-level reranking rather than replacing production retrieval in one step.

retrieval-ranking (retrieval)

Retrieved item/store candidates are reranked with a Qwen reranker and final production rankers, making semantic recall usable within existing ranking constraints.

batch-offline-processing (infrastructure)

Catalog embeddings are refreshed offline and incrementally so evolving menus and products do not force wasteful full-corpus regeneration.

eval-driven-development (evaluation)

DoorDash selected embeddings through retrieval metrics, semantic fidelity, latency, and index-efficiency evaluation instead of choosing by model reputation.

Tradeoffs

Use content-first embeddings as the bootstrap layer

Chose LLM-generated profiles plus off-the-shelf embedding models over pure co-visitation or engagement-derived embeddings.

Rationale: Content-first profiles gave day-zero semantics and cross-vertical coverage where engagement density was sparse.

Adopt compact MRL embeddings

Chose gemini-embedding-001 with 256-dimensional MRL output over larger-dimensional embeddings when index efficiency mattered.

Rationale: The model/dimension choice balanced quality with operational index efficiency for large-scale search retrieval.

Layer item-level EBR and reranking onto store-level retrieval

Chose item profile embeddings plus Qwen reranking over store-level EBR alone.

Rationale: Dish-specific queries needed item-level relevance and query-relevant imagery that store-level retrieval alone could not provide.

Keep consumer representation hybrid

Chose content semantics for entities plus engagement/sequence modeling for users over text-profile embeddings as a universal representation for consumers.

Rationale: Entity identity is expressible in language, but consumer intent is behavioral and situational.

Gotchas (what bit them)

Behavioral embeddings failed to capture semantic intent for sparse commerce sessions.

Root cause: DoorDash users do not produce social-feed-scale engagement streams, and clicks are an indirect proxy for cuisine, product, and occasion meaning.

Resolution: Bootstrap semantic content embeddings from LLM-generated profiles and use downstream retrieval/ranking to adapt them to business objectives.

Input representation mattered more than simply choosing a stronger embedding model.

Root cause: Noisy raw item metadata limited the benefit of better encoders when the semantic profile itself was weak.

Resolution: Generate richer LLM profiles before embedding and evaluate model/data combinations separately.

Consumer profile embeddings break down when preferences are contextual and temporally shifting.

Root cause: A consumer's intent varies by situation, so a single text profile vector averages away distinctions like weekday lunch versus family browsing.

Resolution: Move toward context-conditioned representations that modulate base engagement history by time, location, occasion, and inference context.

Embedding refreshes can become wasteful when catalogs evolve continuously.

Root cause: Menus and product catalogs change, but regenerating every entity embedding daily would create unnecessary offline inference cost.

Resolution: Use incremental Metaflow inference keyed to underlying content changes.

Evaluation

DoorDash evaluated embedding families and profile inputs using Hit Rate@K, nDCG@K, semantic fidelity, systems latency, index efficiency, offline precision, and online search/homepage product metrics.

Metrics: Hit Rate@K · nDCG@K · semantic fidelity · systems latency · index efficiency · null search rate · core search session CVR · homepage order rate · offline precision@10

Outcomes

• LLM-generated profiles materially improved item/store embedding quality.
• Semantic search reduced null search rates and increased conversion.
• Generative homepage carousels increased order rate and offline precision.

The system is deployed across DoorDash search and homepage recommendation surfaces using 256-dimensional embeddings, incremental catalog inference, and production search/ranking integrations.