simple-reranker

A simple ruby app, using elasticsearch and metarank to rerank search results, orchestrated with docker-compose

TL;DR

After downloading and running the docker-compose up command, you should be able to use the postman collection to:

• Set up a movies index in Elasticsearch.
• Search the movies index via the /search endpoint, using the q query parameter, ie http://localhost:4567/search?q=toy.
• Rerank the search request via the /rerank endpoint.
• Send click events to metarank and get personalized reranking results.

Initial Elasticsearch Query for "toy"

GET http://localhost:4567/search?q=toy
[
    {
        "id": 3114,
        "title": "Toy Story 2"
    },
    {
        "id": 78499,
        "title": "Toy Story 3"
    },
    {
        "id": 2253,
        "title": "Toys"
    },
    {
        "id": 1707,
        "title": "Home Alone 3"
    },
    {
        "id": 2953,
        "title": "Home Alone 2: Lost in New York"
    },
    {
        "id": 6936,
        "title": "Elf"
    },
    {
        "id": 2399,
        "title": "Santa Claus: The Movie"
    },
    {
        "id": 90866,
        "title": "Hugo"
    }
]

Initial Reranking Query for "toy"

GET http://localhost:4567/rerank?q=toy
[
    {
        "id": 2953,
        "title": "Home Alone 2: Lost in New York"
    },
    {
        "id": 3114,
        "title": "Toy Story 2"
    },
    {
        "id": 1707,
        "title": "Home Alone 3"
    },
    {
        "id": 78499,
        "title": "Toy Story 3"
    },
    {
        "id": 90866,
        "title": "Hugo"
    },
    {
        "id": 6936,
        "title": "Elf"
    },
    {
        "id": 2253,
        "title": "Toys"
    },
    {
        "id": 2399,
        "title": "Santa Claus: The Movie"
    }
]

Reranking Query for "toy" After Click Event

POST http://localhost:8080/feedback
{
    "event": "interaction",
    "type": "click",
    "id": "id2",
    "ranking": "id1",
    "item": "3114",
    "user": "alice",
    "session": "alice1",
    "timestamp": 1661431896711
}

//then
GET http://localhost:4567/rerank?q=toy
[
    {
        "id": 3114,
        "title": "Toy Story 2"
    },
    {
        "id": 2953,
        "title": "Home Alone 2: Lost in New York"
    },
    {
        "id": 78499,
        "title": "Toy Story 3"
    },
    {
        "id": 1707,
        "title": "Home Alone 3"
    },
    {
        "id": 90866,
        "title": "Hugo"
    },
    {
        "id": 6936,
        "title": "Elf"
    },
    {
        "id": 2253,
        "title": "Toys"
    },
    {
        "id": 2399,
        "title": "Santa Claus: The Movie"
    }
]

Libraries/Tools

• Sinatra
• Elasticsearch (version 8.7.0)
• Metarank
• Docker Compose (tested on Docker Desktop for Mac version 4.12)
• Postman

Data

The movie data used the ranklens dataset from the metarank demo

This dataset is a MovieLens-based crowd-sourced training data for ranking models. It has:

• Movie metadata from TMDB: actors, genres, tags, dates.
• Rankings for 2100 visitors for random movie groups.
• Each visitor interacted with ~5 groups, and marked >1 movies as liked.

Structure

The ruby app uses the Sinatra framework to expose endpoints for setting up an elasticsearch index, searching the index, and reranking the results via metarank.