Neural Search and Conversational AI for Food Delivery and Restaurant Discovery

swiggy · E-commerce · 2023 · source Agentic: yes · promotional bias: med · summary-only (no fetched source) · 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: Food delivery discovery

Swiggy needed food, grocery, restaurant, and support discovery experiences that could handle natural open-ended user intent across a very large catalog.

Constraints:

• The food catalog contains more than 50 million items.
• Users express intent conversationally rather than with precise keywords.
• The approach needed to extend across food delivery, Dineout, Instamart, partner tools, and customer service.
• The available source cache is not usable, so claims rely on the ZenML summary.

Architecture

Swiggy built in-house neural search powered by a fine-tuned LLM for open-ended food discovery, then extended conversational AI patterns across restaurant discovery, grocery shopping, partner support, and customer service.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  natural-language-discovery-interface["Natural Language Discovery Interface"]
  fine-tuned-neural-search-model["Fine-Tuned Neural Search Model"]
  food-and-grocery-catalog["Food and Grocery Catalog"]
  ecosystem-conversational-bots["Ecosystem Conversational Bots"]
  recommendation-output["Recommendation Output"]
  natural-language-discovery-interface --> fine-tuned-neural-search-model
  fine-tuned-neural-search-model --> food-and-grocery-catalog
  food-and-grocery-catalog --> recommendation-output
  ecosystem-conversational-bots --> fine-tuned-neural-search-model

Components

Data flow

Natural user requests are interpreted by the neural search and conversational systems, mapped against catalog or ecosystem inventory, and returned as personalized food, grocery, restaurant, or support recommendations.

Why this architecture

Fine-tuned neural search was used because food discovery queries are open-ended and personalized, not just keyword lookups.

Models used: Fine-tuned Large Language Model · GPT-4 powered customer service chatbot

Scale

• Swiggy's catalog contains more than 50 million items.
  “50 million-plus item catalog.”summary, stated

Alternatives considered (roads not taken)

Keyword-only catalog search

Rejected because: Open-ended user needs like workout meals require semantic intent understanding.

One-off food search feature only

Rejected because: Swiggy extended the conversational approach across restaurant, grocery, support, and customer service experiences.

Patterns

retrieval-ranking (retrieval)

Neural search maps conversational food intent to ranked catalog recommendations.

fine-tuning-adaptation (model-adaptation)

Swiggy fine-tuned an LLM for food-discovery search over its catalog.

tool-use-orchestration (assistant)

Conversational bots are applied across ecosystem tasks including restaurant discovery, grocery shopping, partner support, and customer service.

Tradeoffs

Build in-house neural search

Chose Fine-tuned LLM-powered search over Generic keyword or faceted search.

Rationale: Natural intent and personalization are central to food discovery over a massive catalog.

Use partner model collaboration for customer service

Chose Third-party GPT-4 powered chatbot collaboration over Only the in-house neural search model for all use cases.

Rationale: Customer service chatbot requirements differed from catalog discovery.

Gotchas (what bit them)

Large catalogs create choice overload that keyword search does not solve.

Root cause: Users may know goals or contexts rather than exact item names.

Resolution: Use conversational neural search to interpret open-ended intent.

A discovery pattern must adapt to multiple marketplace surfaces.

Root cause: Food delivery, restaurants, grocery, partner support, and customer service have different workflows.

Resolution: Extend conversational AI across the ecosystem rather than leaving it isolated.

Evaluation

The summary describes deployment breadth but does not provide quantitative retrieval or conversion metrics.

Metrics: No quantitative metrics reported

Outcomes

• Users can issue open-ended natural-language food discovery queries.
• Conversational bots were extended to Dineout, Instamart, partner support, and customer service.
• The approach supports personalized recommendations across a 50 million-plus item catalog.

The catalog contains more than 50 million items.