Ecommerce ops · Production

Picnic enhances grocery search retrieval with LLMs and precomputed embeddings

The problem

Picnic needed to serve a high volume of grocery search queries across multiple countries with different languages and culinary preferences, while handling typos, vague inputs, and the latency demands of real-time as-you-type search.

Workflow diagram · grounded in source
1
Customer submits search query
trigger
“our customers use search to navigate our broad product and recipe assortment, with millions of different search terms being used in the process”
2
LLM generates product description
ai_action
“we dynamically generate descriptions that capture the essence of articles and recipes, transforming search terms into detailed, actionable queries. For this we are using OpenAI's GPT3.5-turbo model”
3
Convert description to embeddings
ai_action
“The output prompt is converted into embeddings using the text-embedding-3-small model from OpenAI”
4
Precompute embeddings for speed
ai_action
“it is easier to precompute 99% of search terms than to set up infrastructure and introduce dependencies that would not allow for milliseconds of latency”
5
OpenSearch retrieval
integration
“At Picnic, we employ OpenSearch to deliver swift search results to our customers”
6
Embedding sanity checks
validation
“we've integrated numerous sanity checks within our pipeline such as verifying if the embeddings are consistent and of the appropriate length”
7
A/B test and iterate
feedback_loop
“online A/B testing, a crucial phase where new features are introduced to a controlled group of users. This testing method allows us to collect valuable data on how real users interact with the changes”
Reported outcome

Picnic implemented LLM-powered prompt-based description generation with precomputed embeddings and OpenSearch retrieval, aiming to improve conversion rates, click-through rates, and customer satisfaction, validated through iterative A/B testing.

Reported metrics
Search terms precomputed99%
Conversion rateimprove conversion rates
Click-through rateenhancing the click-through rate
Customer satisfactionboost customer satisfaction
Reported stack
GPT3.5-turbotext-embedding-3-smallOpenSearch
Source
https://blog.picnic.nl/enhancing-search-retrieval-with-large-language-models-llms-7c3748b26d72
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Picnic implemented LLM-powered prompt-based description generation with precomputed embeddings and OpenSearch retrieval, aiming to improve conversion rates, click-through rates, and customer satisfaction, validated th…

What tools did this team use?

GPT3.5-turbo, text-embedding-3-small, OpenSearch.

What results were reported?

Search terms precomputed: 99%; Conversion rate: improve conversion rates; Click-through rate: enhancing the click-through rate; Customer satisfaction: boost customer satisfaction (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Customer submits search query → LLM generates product description → Convert description to embeddings → Precompute embeddings for speed → OpenSearch retrieval → Embedding sanity checks → A/B test and iterate.

Related ecommerce ops cases