Ecommerce ops · Production

Swiggy's generative AI year in review: catalog enrichment, review summarization, neural search, and restaurant partner RAG support

The problem

Swiggy needed to improve food catalog coverage through image and text description generation, reduce customer decision fatigue during ordering, and streamline dense FAQ navigation for restaurant partners, while managing generative AI risks including hallucination, latency, and data governance.

Workflow diagram · grounded in source
1
Food image generation
ai_action
“we integrated the Stable Diffusion pipeline and customized it for Swiggy's needs. We explored three approaches for food image generation: Text2Img, Img2Img, Image blending”
2
LoRA fine-tuning for Indian cuisine
ai_action
“The StableDiffusion (SD) model (v1.5) was fine-tuned using LoRA to enhance the generation quality specifically for Indian dishes such as dosa, curry, Indian breads, biryani, and others”
3
Dish description text generation
ai_action
“we deployed a customized generative AI pipeline for increasing the coverage of item descriptions for the dishes on the catalog”
4
Human review of descriptions
human_review
“A human agent in the loop, sanity checks the descriptions and provides feedback for improvement if necessary”
5
Review summarization via GPT4
ai_action
“We leveraged GPT4 with customized prompts for generating review summaries and implemented an internal evaluation metric to establish the quality/customer acceptability of the reviews”
6
A/B test evaluation
feedback_loop
“In an A/B test involving over 2K restaurants, we observed improvements in funnel metrics, and reductions in cancellations and claims, attributed to enhanced expectation management”
7
Restaurant partner RAG query handling
ai_action
“an LLM powered bot was developed to allow users to input their queries directly, fetching relevant answers without the need for manual search”
8
WhatsApp response delivery
output
“allows responses in both Hindi and English through WhatsApp, addressing a wide range of questions based on standard operating procedure (SOP) documents”
Reported outcome

Swiggy deployed generative AI across catalog image and text enrichment, review summarization, an AI content flywheel, neural search, and a restaurant partner RAG bot.
A review summarization A/B test across over 2K restaurants showed improvements in funnel metrics and reductions in cancellations and claims. The restaurant partner bot was deployed to a subset of partners with promising initial results.

Reported metrics
Average menu browsing time~10–20 mins
restaurants in A/B testover 2K
Funnel metricsimprovements in funnel metrics
Cancellationsreductions in cancellations
Show all 5 reported metrics
average menu browsing time~10–20 mins
restaurants in A/B testover 2K
funnel metricsimprovements in funnel metrics
cancellationsreductions in cancellations
claimsreductions in claims
Reported stack
Stable DiffusionLoRAGPT4OpenAI APILLMRAGNLUPythonvector databasesWhatsApp
Source
https://bytes.swiggy.com/reflecting-on-a-year-of-generative-ai-at-swiggy-a-brief-review-of-achievements-learnings-and-13a9671dc624
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Swiggy deployed generative AI across catalog image and text enrichment, review summarization, an AI content flywheel, neural search, and a restaurant partner RAG bot.

What tools did this team use?

Stable Diffusion, LoRA, GPT4, OpenAI API, LLM, RAG, NLU, Python, vector databases, WhatsApp.

What results were reported?

Average menu browsing time: ~10–20 mins; restaurants in A/B test: over 2K; Funnel metrics: improvements in funnel metrics; Cancellations: reductions in cancellations (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Food image generation → LoRA fine-tuning for Indian cuisine → Dish description text generation → Human review of descriptions → Review summarization via GPT4 → A/B test evaluation → Restaurant partner RAG query handling → WhatsApp response delivery.

Related ecommerce ops cases