Search Types

Local, online, and offline search.

[1] Straightforward results.

You can try Palet for yourself by following this link. Search queries at the moment are limited to Agentic Search. Keep in mind that every query you submit will generate real results. Meaning that dozens of live phone calls will be made to real businesses.

At the moment, Palet is limited to a very early version of Agentic Search. We built it in 10 days. However, we are working on making it better and shipping Deep and Local Search features by early February 2025. GitHub repo can be found here.

Agentic Search

One of the limitations of traditional search engines is that they can’t generate results outside of what can been indexed on the World Wide Web. A good example of this is if you're looking to check whether an ice cream shop has a particular flavor of vegan ice cream. Traditional search engines wouldn't be able to return a final result, leaving it to the user to call shops to check for availability. This includes LLM-based search engines like ChatGPT and Perplexity.

[2] Perplexity is a case example of an LLM-based search engine that can't complete a search query end-to-end.

Palet uses a search agent that employs an exhaustive search strategy to brute-force a result. For any search, a reasoning model first spends time understanding the user query. Using the ice cream example, the reasoning models determine that a phone call needs to be made to all nearby ice cream shops in order to obtain a final result. And so it instantiates an ensemble system that plans and makes dozens (up to 2,100 concurrent calls at the moment) of phone calls in parallel to synthesize an answer for the user.

[3] Palet is able to call ice cream shops to confirm availability and generate end-to-end results.

In the future, we plan on expanding agentic search for the model to be able to write to public forums such as Twitter, StackOverflow, Reddit, and more. And for it to be able to sync your contacts for personal requests, such as reminding a family member about something.

Deep Search

Search engines return ranked results full of sponsored links and search engine optimized content. Ranking depends on relevance, backlinks, user behavior, and other blind signals to determine what makes it to the front page. Whereas most LLM-based search simply summarize those same ranked results. And usually, they're limited to summarizing the links on the first page—limited by their context window.

[4] No one ever reads past the first page of Google Search results.

Palet's search iteratively refines candidate results by searching, determining relevance, and re-searching, much like a human would. This loop can run thousands of times, culminating in a curated, context-aware set of results rather than a list ranked by blind signals. Instead of summarizing content, Palet presents results much like a Google search would. It turns days of combing through traditional search results into minutes. While still giving the user what they want...results, not summaries.

[5] Palet curates results and then ranks them for you. Allowing you to read search results for yourself, rather than a summary.

Local Search

Open Context Protocol is an open standard for establishing secure, bidirectional connections between data sources and client applications—like Palet. Its architecture parallels the AT Protocol (Bluesky) but is purpose-built to let AI clients interact seamlessly with your data. By connecting an application (e.g., iMessage, calendar) via OCP, its contents become searchable and accessible to Palet's models. Most search engine's knowledge graphs are proprietary and closed source. We’re committed to building OCP as an open-source project and ecosystem. It's meant to serve as a standard filesystem for AI applications (agentic or not), similar to what ActivityPub (Threads, Mastadon) and ATProtocol (Bluesky) are for large-scale distributed social applications.