"A new way to save, share and search"—this is our mandate. At its core, this means we're evaluating existing paradigms of saving, sharing, and search, and attempting to evolve and/or revolutionise them. Happily, this allows us to ask deep and interesting questions, such as:
What's the best way to represent a search space?
Let's define "representation of a search space" as what a user sees after they've provided a query and received a response. Now for an assertion: we're still in the earliest generations of search space representation. We're still bound to 2D digital interfaces constrained by rectangular glass planes; alternatives are difficult to conceive, prototype, bring to market, and scale to consumer-level adoption. A sustainable shift forward from our entrenched and effective defaults requires unusual courage and pioneering spirit, as well as a lot of capital and a long time horizon. So we're stuck in a valley.
Search engines—think Google, product search on Amazon, even Exa's embeddings-based search—provide simple formats like lists and grids, accompanied by sort and filter tools. The current crop of answer engines—think interactions with ChatGPT, responses via Perplexity or Elicit—default to single threaded chat UXs. As Maggie Appleton says in Language Model Sketchbook, or Why I Hate Chatbots, this is:
...only the obvious tip of the iceberg when it comes to exploring how we might interact with these strange new language model agents we've grown inside a neural net.
Outside of these baselines, what's possible? How can a search space—what a user sees after they've provided a query and received a response—be represented? Let's take a look...
List: A one-dimensional (usually vertically oriented) sequence of results with basic information, designed to simplify comparison and facilitate quick click-throughs or navigation to a more detailed page/resource.
Grid: A two-dimensional arrangement of search results, often used for visual content like images or products. Grids facilitate rapid scanning and comparison by placing items in rows and columns with contextualising details, softening an info-dense experience.
Table: A structured format presenting data in rows and columns, ideal for comparing multiple attributes of search results. Tables enable detailed sorting and comparison, making them suitable for datasets with rich metadata.
Tree: A hierarchical representation of search results, showing parent-child relationships between items. This format helps users explore nested data structures, like file systems or organisational hierarchies, by drilling down into specific categories.
Cluster or heatmap: Groups similar search results together, highlighting patterns or themes within the data. Heatmaps add a layer of visual intensity (using colour gradients) to indicate areas of higher relevance or interest, making dense data more interpretable.
Graph: A network of nodes (representing entities) connected by edges (representing relationships). Graphs are powerful for visualising complex relationships and interdependencies among search results, such as social networks or citation networks.
Sequence or state: Represents search results in a specific order, often over time or stages, including branching paths. This is useful for time-series data or processes where understanding the progression of events or conditions—and their various possible outcomes—is key.
Ontology: A structured framework defining a set of domain-specific concepts and their relationships. Ontologies offer a rich, contextualised view of search results, enabling sophisticated querying and reasoning about the entities and their interactions.
Vector or hyperplane: Vectors represent search results as points in a multi-dimensional space; hyperplanes divide this space to create different categories, such as quadrants in 2D diagrams, regions in 3D diagrams, or segments in higher-dimensional spaces. This helps users to visually and mathematically organise and investigate data.
Spatial: Uses maps to present search results, displaying data points based on physical or imaginative locations. This format is vital for real-world geographical searches as well as speculative or designed layouts, providing a visual context that aids in understanding spatial relationships.
Mediator: An abstract interface that simplifies the complexity of the search space by collapsing it to a single touchpoint, often integrating multiple visualisations or aggregating data from various sources. Mediators serve as semi-autonomous go-betweens reminiscent of archetypical butlers, assistants, servants and service providers.
At this point, you may be wondering why we've bothered enumerating the ways to represent a search space. What's the point?
Earlier, we mentioned that we're stuck in a valley when it comes to search space representation and user experience. One way to contribute to our escape from it is through a simple recognition of the reality around us and the terrain we occupy. By defining the possible, the plausible and the probable, we also define the inverse—what's impossible, implausible, improbable. Somewhere between those states—between the "im" and the "p..."—lurks opportunity, and it's our mission to find it.
But this isn't a zero sum game; we want you to start looking, too. So take our search space representations as a foundation and help us all find a better way forward.
Matthew McDowell-Sweet
"Search engines—think Google, product search on Amazon, even Exa's embeddings-based search—provide simple formats like lists and grids, accompanied by sort and filter tools. The current crop of answer engines—think interactions with ChatGPT, responses via Perplexity or Elicit—default to single threaded chat UXs. As Maggie Appleton says in Language Model Sketchbook, or Why I Hate Chatbots, this is: '...only the obvious tip of the iceberg when it comes to exploring how we might interact with these strange new language model agents we've grown inside a neural net.' Outside of these baselines, what's possible?" https://paragraph.xyz/@subset/representing-search-space