AisleFind

Problem

Ever gone grocery shopping and wanted to find a “low-sugar ketchup” or “nut-free strawberry cereal”, ideally, but it was just too tedious to look for?

Many of us have nutritional preferences or dietary restrictions, but the thought of turning over hundreds of products to compare teeny-tiny nutrition labels is enough to make anyone cry.

To save time, we have no choice but to rely on marketing claims in the packaging, and just hope they’re not totally misleading. Even so, this isn’t always an option for everyone, especially those with allergies or strict dietary restrictions.

AisleFind

AisleFind is a recommendation system for grocery food products.

Just tell the AI chatbot what kinds of food products you’re looking for (e.g. “low-fat peanut butter with grape jam”), and AisleFind will semantically search a database of 200k food products to recommend products for your needs.

Claude gets a question and decides to use the "Semantic food search" tool to gather information

Get concise, explained recommendations

Continue chatting about options...

...until you get the answer you need

If you didn’t get what you wanted the first time, feel free to ask follow-up questions, or ask AisleFind to compare products.

Development notes

• Currently, AisleFind is a prototype using Claude Desktop for the UI and access to LLMs, with MCP servers interacting with VoyageAI to embed queries and a Postgres database that stores all the data for each food product (like brands, ingredients and images) as well as embeddings.

• To test with more people, the next step would be to deploy a web-accessible UI. I'm exploring LibreChat since it comes with auth and rate-limiting out-of-the-box.

• I used Open Food Facts (OFF) as my data source as it was the best single source of all the key data, plus product face images. The original dataset numbers 4.2 million food products. I filtered for English-language products with ingredient and nutrition information and product face images - only 200k products remained. Still, this is a good enough number to start.

• Being community-contributed, OFF’s data could contain errors. We can improve confidence in our dataset by cross-checking it with government-provided sources like USDA’s FoodData Central (Branded Foods) data. Identifying product codes in our data that are valid Global Trade Item Numbers (GTIN) could also help retail stores utilise the data for their own applications or indicate availability.

Project by

syan

thanks and credits to Open Food Facts and its contributors, the former FoodFacts team, Kenny Eliason and a lot of YouTubers (: