cs.AIMar 2, 2026

Agentic Multi-Source Grounding for Enhanced Query Intent Understanding: A DoorDash Case Study

Authors: Emmanuel Aboah Boateng, Kyle MacDonald, Akshad Viswanathan, Sudeep Das

Organizations: DoorDash, Inc. San Francisco, California, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Accurately mapping user queries to business categories is a fundamental Information Retrieval challenge for multi-category marketplaces, where context-sparse queries such as "Wildflower" exhibit intent ambiguity, simultaneously denoting a restaurant chain, a retail product, and a floral item. Traditional classifiers force a winner-takes-all assignment, while general-purpose LLMs hallucinate unavailable inventory. We introduce an Agentic Multi-Source Grounded system that addresses both failure modes by grounding LLM inference in (i) a staged catalog entity retrieval pipeline and (ii) an agentic web-search tool invoked autonomously for cold-start queries. Rather than predicting a single label, the model emits an ordered multi-intent set, resolved by a configurable disambiguation layer that applies deterministic business policies and is designed for extensibility to personalization signals. This decoupled design generalizes across domains, allowing any marketplace to supply its own grounding sources and resolution rules without modifying the core architecture. Evaluated on DoorDash's multi-vertical search platform, the system achieves +10.9pp over the ungrounded LLM baseline and +4.6pp over the legacy production system. On long-tail queries, incremental ablations attribute +8.3pp to catalog grounding, +3.2pp to agentic web search grounding, and +1.5pp to dual intent disambiguation, yielding 90.7% accuracy (+13.0pp over baseline). The system is deployed in production, serving over 95% of daily search impressions, and establishes a generalizable paradigm for applications requiring foundation models grounded in proprietary context and real-time web knowledge to resolve ambiguous, context-sparse decision problems at scale.

Figures & tables

Figure 1. Intent Ambiguity for the query “Wildflower.” (Left) The Restaurant category surfaces the fictional Wildflower Bites chain. (Right) The Retail category surfaces literal flower products. Our system captures both interpretations to enable context-aware routing. Side-by-side search results for the query Wildflower. The left side shows a restaurant interpretation represented by the fictional Wildflower Bites chain, and the right side shows retail flower products, illustrating why the query has multiple valid intents.
Figure 2. System Architecture Overview. The pipeline illustrates the multi-source evidence retrieval process (steps 2–4, 6), the dual-intent reasoning engine (step 5), and the pluggable disambiguation layer (step 7) that populates the production cache. Architecture diagram for the query intent system. A user query flows through catalog retrieval and optional external web search, both evidence sources are injected into an LLM that predicts primary and secondary intents, and a downstream disambiguation layer resolves the final category before writing results into a production cache.
QueryBaselineEntity ContextSearch ContextDI Output (p/s)Resolved
“wildflower”FlowerWildflower Bites (Restaurant)Popular restaurant chain in AZRestaurant / FlowerRestaurant
“better chew”Pet ProductBetter Chew Farms (Plant-Based)Plant-based frozen food brandGrocery / DishGrocery
“450 north”Restaurant450 North Craft Ales (Beer)Craft brewery, not a restaurantAlcohol / Retail StoreAlcohol
Table 1. Dual-intent resolution on ambiguous tail queries. Baseline: ungrounded Gemini-2.5-flash.
BenchmarkGPT 4o MiniGPT 4oGemini BaseLegacyDI (Ours)
Branded*77.9%87.5%88.4%82.3%94.5%
Retail86.5%93.6%90.5%91.2%96.3%
Tail (Synthetic)82.6%87.1%86.4%83.4%92.4%
SOT (Overall)79.8%83.2%83.1%89.4%94.0%
Table 2. Accuracy Comparison across Benchmarks.
Figure 3. SOT Accuracy by Segment. Bar chart comparing GPT-4o-mini, GPT-4o, Gemini Base, the legacy system, and the proposed dual-intent system across Head, Torso, and Tail traffic. The dual-intent system is best on all segments, with the largest gap on Tail queries where it reaches 90.7 percent versus 82.2 percent for the legacy system and 77.7 percent for Gemini Base.
Figure 4. Component Ablation on SOT Tail ( N=4,993N{=}4{,}993 ). Cumulative +13.0pp lift over the ungrounded baseline. Bar chart showing incremental gains from each component on the SOT Tail benchmark. Accuracy rises from 77.7 percent for the ungrounded baseline to 86.0 percent with catalog grounding, 89.2 percent after adding agentic search, and 90.7 percent for the full system with dual-intent disambiguation.
More in this topic

E-Commerce Search

E-commerce search studies how to match shoppers’ queries, preferences, and product signals with useful items in online marketplaces.

See all 42 papers

Explore similar work

CardsList
Jun 17, 2026

Decoupling Search from Reasoning: A Vendor-Agnostic Grounding Architecture for LLM Agents

Production LLM agents increasingly depend on real-time search, yet native search grounding bundles retrieval policy, provider choice, evidence injection, cost, latency, and generation behavior behind a single model-provider boundary. This coupling makes grounding hard to inspect, tune, reuse, or port, and can trigger Search-Induced Verbosity that breaks strict output contracts. We present Decoupled Search Grounding (DSG), a vendor-agnostic boundary that moves grounding outside the reasoning model through an MCP-compatible gateway, exposing provider routing, source-aware context rendering, configured fallback, retrieval-depth control, and exact plus semantic caching as first-class controls. Across five frontier models on SimpleQA, FreshQA, and HotpotQA, native search leads on recency-sensitive FreshQA, but DSG exposes a stronger frontier when control matters: on SimpleQA it nearly matches native accuracy (86.1% vs. 87.7%) at 91% lower search cost, preserves concise answer contracts, and reaches a 99.4% warm-cache hit rate with 68% lower latency. Deployed as a shared production grounding layer for large-scale agentic workloads with interchangeable models, DSG matches native-search accuracy on an e-commerce query-understanding (QIU) workload while cutting search cost by over 98%. Real-time grounding is best treated as an optimizable interface boundary, not a fixed model feature.

Emmanuel Aboah Boateng, Kyle MacDonald, Amardeep Kumar +2

DoorDash, Inc.

LLM GroundingTool-Augmented Language Model AgentsAgentic SearchAgentic RAG
AskPDF
Aug 16, 2026

Dear Algo: A Precision-First Agentic Intent Layer for Unified Search and Recommendation

Search and recommendation serve a shared discovery objective but encode intent differently. We study this boundary through Dear Algo on Threads, a deployed product where open-ended requests such as more NBA news or less politics steer subsequent feed recommendations rather than return a one-shot result list. Its agentic intent layer compiles explicit, inferred, negative, and compound intent into a grounded executable plan, then invokes conventional retrieval and optional semantic or multimodal reranking. The layer shares an intent-to-retrieval contract without requiring one model or serving path across search-like and recommendation-like modes. We evaluate Dear Algo under a precision-first objective. In a blinded audit of 300 public request-item pairs (296 evaluable), a strict categorical LLM-as-a-judge gate achieved 94.4% exact-Relevant precision [88.8%, 98.9%]. Across 72 normalized request clusters, the full configuration produced 7.73 judge-qualified candidates per 20 slots versus 6.61 for an LLM-derived-query baseline, a gain of 1.11 [0.12, 2.12]. In a candidate-randomized serving-path study restricted to the reranker path's first 72 eligible hours, the user-weighted judge-Irrelevant share among judged admissions was 2.80% versus 4.78% off (-1.97 points [-3.02, -0.94]), while Exact-Relevant share was 2.24 points higher [0.08, 4.41]. Together, these studies show how explicit natural-language intent can be carried into feed recommendation under a precision-first evaluation framework

Rui Wang, Jiazhou Wang, Zheng Wei +16

Meta Platforms Inc. USA · Google DeepMind, USA

Language Model SteeringRecommender SystemsLarge Language Model-Based RecommendationUser Preference Modeling
AskPDF
Oct 1, 2026

RPTune: Learned Context Curation for LLM Catalog Search

For small merchant businesses (SMBs) whose catalogs fit within a long-context LLM, full-catalog prompting offers a compelling alternative to multi-stage retrieval designed primarily for large marketplaces with millions of items. However, fitting the full catalog into the context window does not ensure that the model can use it effectively, since LLMs do not exploit long contexts uniformly. We therefore study in-context catalog search through two complementary questions: (1) how to curate and present catalogs to the LLM, and (2) how to adapt the LLM for product selection on curated contexts. We propose RPTune, an end-to-end framework that couples learned catalog curation with LLM post-training using automatically generated, catalog-grounded supervision. An encoder-reorganizer curator orders and prunes products guided by downstream LLM feedback, while the resulting curated catalogs in turn improve the effectiveness of LLM post-training with a context-relative reward. We evaluate RPTune on 7 real merchants spanning distinct retail verticals, using 100 complex conversational queries per merchant. RPTune consistently improves search accuracy across both proprietary and open-weight LLMs, with context curation yielding gains of up to 31.4 percentage points and post-training adding a further 10.3 points on average.

Chuxuan Hu, Hejie Cui, Norman Huang +3

Google · University of Illinois Urbana-Champaign

In-Context RetrievalLLM Fine-TuningLarge Language Model-Based Recommendation
AskPDF