Gist

Platforms

iOS, macOS, tvOS, visionOS, Android, Linux, Windows, Browser, Node

Languages

101

Weights

v2.2.0

Demo

https://desertant.com/models/gist/

Install

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.5.0")

implementation("ai.desertant:gist:3.5.0")

npm i @desert-ant-labs/gist @litertjs/core   # browser
npm i @desert-ant-labs/gist                  # Node, prebuilt native core

Usage

import Gist

let gist = Gist()
let topics = try await gist.classify("How to start a podcast with just your iPhone")
// [Topic(slug: "technology", name: "Technology & Software", score: 0.93), ...]

let all = try await gist.scores(of: text)        // [String: Double], 36 entries

let posts = titles.map { PostTopics(topics: ..., timestampMillis: ...) }
let channel = channelTopics(posts, options: RollupOptions(topN: 5))
// [ChannelTopic(slug: "technology", share: 0.41, postCount: 12), ...]

import ai.desertant.gist.Gist

Gist(context).use { gist ->
    val topics = gist.classify("How to start a podcast with just your iPhone")
    // List<Topic>: slug, name, score
    val all = gist.scores(text)                  // Map<String, Double>
}

import { Gist } from "@desert-ant-labs/gist";           // browser
// import { Gist } from "@desert-ant-labs/gist/native"; // server-side Node

const gist = await Gist.load();
const topics = await gist.classify("How to start a podcast with just your iPhone");
// [{ slug: "technology", name: "Technology & Software", score: 0.93 }, ...]
gist.dispose();

import { Gist, channelTopics } from "@desert-ant-labs/gist";

channelTopics(posts, { topN: 5 });
// [{ slug: "technology", share: 0.41, postCount: 12 }, ...]

let gist = Gist(variant: .english)

Files

Inputs and outputs

• Input: a plain text string (title, or title + description). Best on short text like posts, titles, and descriptions.
• Output: a probability over the 36 topics (features [1, 8448] → topic_probs [1, 36]); take the top-k above the threshold in gist_config.json. Optimized for multi-label use: an item's 2-3 topics, optionally aggregated across a collection.

Topics and standard taxonomy

The 36 topics map to two industry-standard taxonomies so gist output can be rolled up or joined into existing systems: IAB Content Taxonomy 2.2 (with each node's stable integer ID) and Apple Podcasts categories. The full, machine-readable crosswalk ships in this repo as taxonomy_crosswalk.json (e.g. law → IAB 383 News & Politics › Law, crafts-hobbies → IAB 248 Arts and Crafts, finance → IAB 391 Personal Finance).

Five topics have no dedicated IAB 2.2 node and are flagged as gist extensions (society-culture, creator-economy, outdoors-nature map to a nearest parent; history and self-improvement have no IAB node); film-tv is a roll-up of IAB Movies + Television.

Languages

Topic tagging covers 101 languages. A diverse 15-language spot check (across Latin, Cyrillic, Arabic, CJK, Devanagari, Hebrew, Thai, and Greek scripts) gives 88% top-3, with CJK, Arabic, and Cyrillic scripts matching or beating the Latin ones.

Model variants

Two builds of the same 36-topic model live in this repo:

The English build is the same model with a smaller embedding and tokenizer, so it is topic-identical to the multilingual model on English input. It does not cover non-Latin scripts (CJK, Arabic, Cyrillic, …); use it only when the input is reliably English/Latin. The Swift SDK selects it with Gist(variant: .english). The JS and Kotlin SDKs currently load the multilingual build only: variant selection has to cross the shared native ABI, which has no slot for it yet.

Evaluation

Recall on a held-out set of 572 human-labeled real posts (36 topics), zero-shot for the LLMs and zero-shot classifiers. Embedding classifiers get a light logistic head; recall@3 is the product metric (downstream aggregation consumes the top few topics).

gist is tied on recall@3 with the best small models, at a fraction of the size and one on-device pass, and it beats every zero-shot classifier decisively (they never learned the taxonomy or the distribution). Only a 7B cloud LLM clearly leads on recall@1.