> Explore Qdrant's agent skills catalog at https://skills.qdrant.tech/
> Search the documentation at https://skills.qdrant.tech/search?query=your+query+here
> Use this file to discover all available pages: https://qdrant.tech/llms.txt
> Editor's note: This 2023 post was edited for length and clarity. Read the [original version](https://github.com/qdrant/landing_page/blob/bb7f15b97237c97748fdbeea45499e2fcaba2377/qdrant-landing/content/articles/geo-polygon-filter-gsoc.md).

I'm Zein Wen, and I worked on polygon filtering with Arnaud Gourlay during Google Summer of Code 2023. Restaurant recommendations often need a geographic boundary as well as a similarity score. A circle or rectangle cannot describe every neighborhood, so my project added polygon filters to Qdrant.

This post records my 2023 contribution. For current usage, see the [geo polygon filtering documentation](https://qdrant.tech/documentation/search/filtering/index.md#geo-polygon).

## Finding Points Inside a Boundary

Before this project, Qdrant supported radius and rectangle filters. Polygon filtering let users describe irregular areas and combine them with vector search.

<figure><img src="/blog/geo-polygon-filter-gsoc/geo-filter-example.png"
    alt="Map of London with locations inside an irregular search boundary"><figcaption>
      <p>A geographic search area. Source: <a href="https://traveltime.com/blog/map-postcode-data-catchment-area" target="_blank" rel="noopener nofollow">TravelTime</a>.</p>
    </figcaption>
</figure>


The geographic index uses geohashes to group locations into rectangular cells. During a query, it finds cells that could overlap the polygon, then checks candidate locations against the polygon itself. This avoids testing every stored location.

<figure><img src="/blog/geo-polygon-filter-gsoc/geo-index.svg"
    alt="Stored locations are grouped by geohash cell, overlapping cells supply candidates, and a polygon check returns matching locations"><figcaption>
      <p>The index narrows the candidates before checking the polygon boundary.</p>
    </figcaption>
</figure>


Two geometry operations needed careful testing: checking whether a polygon intersects a rectangle and whether a point lies inside a polygon. The Rust `geo` library provided these operations, but we still needed to understand their behavior and verify edge cases.

I explored winding-number and ray-casting algorithms and used visual tests to compare results. That work helped me learn an unfamiliar part of the codebase through small, testable questions.

<figure><img src="/blog/geo-polygon-filter-gsoc/geo-computation-testing.png"
    alt="Six polygon test plots showing different boundaries and point locations"><figcaption>
      <p>Geometry test cases from the project.</p>
    </figcaption>
</figure>


## Keeping the API Consistent

We considered [GeoJSON](https://geojson.org/) for the interface. Its coordinate representation differed from Qdrant's existing radius and rectangle filters, so we kept the polygon interface consistent with those filters.

We also considered a separate multi-polygon filter. Combining polygon conditions already covered that use case, just as users could combine circles or rectangles. Adding another filter type would have increased the API's complexity without adding a necessary operation.

This was a useful lesson in API design. A familiar standard can help users, but consistency with the surrounding API matters too.

## Learning Through Open Source

This was my first opportunity to write Rust for a production project. Arnaud and the other Qdrant engineers helped me work through unfamiliar code, compare alternatives, and explain my decisions during review.

I learned to keep the user's task in mind when designing an interface. I also gained confidence in discussing different approaches and asking questions before committing to a design.

Thank you to everyone who reviewed the work and helped me contribute. To try polygon filtering, follow the [current documentation](https://qdrant.tech/documentation/search/filtering/index.md#geo-polygon) with a [Qdrant Cloud cluster](https://cloud.qdrant.io/) or a [local deployment](https://qdrant.tech/documentation/quickstart/index.md).