Glossary
Embed Model (Content personalization, discovery, moderation): Pre-built AI models by Embed designed to offer personalized content, facilitate discovery, and ensure moderation. These models use rich datasets to provide accurate and relevant user experiences.
Content Personalization: Pre-built Embed AI model to predict user interactions and tailor content feeds, such as social media posts, articles, and NFTs, based on individual preferences to increase engagement, retention, and revenue.
Discovery: Pre-built Embed AI model that helps users find desired content through chat-based interfaces and improved search experiences, enhancing user satisfaction.
Moderation: Pre-built Embed AI model that identifies inappropriate, spammy, or engagement-farming content, ensuring a safer online environment.
AI Label: Tags or classifications assigned by AI models to different types of content to facilitate organization, discovery, and moderation.
Dataset: Collections of data used to train AI models. In the context of Embed, these include on-chain and off-chain data, multimedia, and multilingual datasets to improve the accuracy and relevance of AI models.
Endpoint: A specific URL path on Embed's API that provides a distinct capability — for example, /deploy/serve for serving feeds or /notifications for polling smart money signals.
Handle: A unique identifier for users on social platforms integrated with Embed's AI models, often used in conjunction with user profiles for personalized content delivery.
FID (Farcaster ID): A unique identifier within the Farcaster protocol, used to manage user identities and interactions within decentralized social networks.
Wallet Address: A unique identifier for blockchain-based wallets, used within Embed's ecosystem to link user activities and preferences with their on-chain data.
Embed ID: A unique identifier assigned to each user within Embed's ecosystem. This ID is used by Embed models on social platforms to manage user identities. It serves as the highest hierarchy data mapping tool, superseding other identifiers such as FID (Farcaster ID), wallet addresses, and handles.
AI Model Scoring: The process of evaluating AI model performance based on various metrics to determine its accuracy and effectiveness in content personalization, discovery, and moderation models.
Custom Scores: User-defined scoring metrics used to tailor AI model outputs to specific needs or preferences, enhancing the customization of AI-driven content and interactions.
Model Training: The process of teaching an AI model to make accurate predictions by feeding it large amounts of data and adjusting its parameters.
Data Preprocessing: The steps taken to clean and prepare data for use in training AI models. This includes tasks like removing duplicates, handling missing values, and normalizing data.
Inference: The stage where the trained AI model makes predictions or decisions based on new, unseen data.
API Key: A unique code provided to developers to access Embed's APIs. It is used for authenticating and authorizing API requests, ensuring that only authorized users and applications can interact with Embed's services and endpoints.
API Endpoint: A specific URL where an API can be accessed by developers. Endpoints are used to interact with different functionalities of the Embed platform, such as content personalization, discovery, and moderation.
Algorithm & Feed Terms
Feed: A personalized, ranked list of content items (posts, coins, markets) recommended to a user based on their preferences, interaction history, and configured filters.
Config ID (config_id): A unique identifier for a saved feed configuration. Config IDs are created in the Console or via the Deployments API and used with POST /deploy/serve to serve feeds in production.
Feed Algorithm: A JavaScript function written using the Algorithms DSL that defines a feed pipeline — search, features, scoring, and ranking. Algorithms are saved via the Console or POST /deploy/algos and referenced by feed configs.
Candidate Generation (Search): The first stage of the feed pipeline where items are retrieved from Elasticsearch indices using filters, boosting, or semantic similarity. See the [Search Guide](../Building Algorithms/search).
Features & Scoring: The second stage where candidates are enriched with ML-computed signals (topic relevance, user affinity) and optionally reranked by trained models. See the Features Guide.
Ranking: The third stage where signals are combined into a final ordering using sort, linear, or mix methods, with optional diversity controls. See the Ranking Guide.
Feed Config: A deployment configuration that bundles one or more algorithms with weights, fallback chains, cache TTLs, and exclusion rules. See the Feed Configs Guide.
Cold Start: A strategy for handling users with limited or no interaction history. When a user is new or inactive, a cold start feed configuration is used to provide quality content until enough interaction data is available for personalization.
Fallback Feeds: Alternative feeds specified to show when the main feed runs out of content. Fallback feeds ensure users always have content to see, even when the primary feed is exhausted.
Promotion: Configuration for mixing promoted or sponsored content into the organic feed. Supports two types: 'feed' (blend another feed) or 'items' (insert specific posts at defined positions).
Social Proof: Indication that users in the viewer's network have interacted with a piece of content. A post has social proof if at least one user that the viewing user follows has interacted with it (liked, commented, or shared).
Impression Count: The number of posts that will be considered as "seen" by the user in a given API call. When set, the top n items are marked as viewed and will not appear in subsequent API responses for that user, helping maintain feed freshness.
Data Source Terms
Onchain Graphs: Data recorded on blockchain networks that is transparent, verifiable, and permanently stored. This includes user identities, interactions, and content metadata recorded on the blockchain.
Offchain Graphs: Data from social platforms and applications that is not recorded on blockchain networks. This provides rich context about user behavior, content engagement, and social relationships.
Agents: User profiles or wallet addresses that interact with onchain protocols. Agents can be standard social media profiles or profiles associated with wallet addresses (e.g., Farcaster IDs, Zora creators, Polymarket traders).
Intent: User actions and interactions recorded onchain. These include both user-to-user and user-to-item interactions that indicate preferences, interests, and engagement patterns (e.g., mints, purchases, trades, bets, follows).
Assets: Onchain content or digital items that users interact with. These can be user-generated content (text, images, audio, video posts) or onchain tokens, NFTs, or market contracts (e.g., Farcaster casts, Zora coins, Polymarket markets).
Impressions: Tracking of when users view or are exposed to content, even if they don't explicitly interact with it. This data helps understand what content users have seen and prevents showing duplicate content.
Metadata: Rich information about content and users generated through AI analysis and platform data. This includes AI labels, embeddings, text analysis, user profiles, and engagement patterns.
Catalog: The comprehensive inventory of all available content and users in the system. It's the complete database that feeds use to generate candidates for recommendation.
Embeddings: Vector representations of content (media, text, users) that capture semantic meaning and enable similarity calculations. Embeddings are used for content analysis, user profiling, and recommendation matching.
Scoring and Ranking Terms
Scoring Algorithm: The method used to rank and prioritize content in a feed. Different algorithms emphasize different signals (interest, affinity, trending, popularity) to create personalized recommendations.
Balanced Feed: A scoring algorithm that combines interest-based and affinity-based signals to create a balanced mix of personalized content. This is the default recommendation algorithm.
Interest Score: A metric that reflects how likely a user is to interact with content based on their past interaction history and inferred preferences.
Affinity Score: A metric that reflects how likely a user is to interact with content based on their social graph and following relationships.
Trending Score: A metric that reflects how quickly content is gaining engagement and attention in the network.
Popular Score: A metric that reflects the overall popularity and engagement level of content across all users.
Time Decay: A mechanism that reduces the score of older content over time, helping keep feeds fresh by prioritizing recent content while still allowing high-quality older content to surface.
Diversity: Controls that ensure varied content in the feed, preventing over-representation of individual authors or topics. This includes limits on posts per author and minimum distance between posts from the same author.
Weighted Scoring: A custom ranking approach where specific features (interest scores, trending scores, affinity scores, etc.) are assigned weights to build a personalized ranking algorithm.
Platform-Specific Terms
Cast: A post on the Farcaster protocol. Casts can contain text, images, videos, embeds, and references to onchain content.
Zora Coin: An NFT or token created on the Zora platform. Zora coins have onchain metadata, ownership history, trading data, and market performance metrics.
Prediction Market: A market contract on Polymarket where users can bet on the outcome of events. Markets have onchain resolution conditions, trading history, and performance data.
Channel: A community or topic-based grouping on Farcaster. Channels allow users to organize content around specific themes or communities (e.g., "Founders" channel).
App FID: The Farcaster ID of an app or client that created content. Used to filter content by source application (e.g., Warpcast, other Farcaster clients).
Polymarket Market: A prediction market contract on Polymarket where users trade outcome shares. Markets have onchain resolution conditions, trading volumes, and probability curves.
Kalshi Event: A prediction market event on Kalshi (CFTC-regulated) where users trade binary outcome contracts on real-world events.
Convergence Signal: A smart money signal fired when multiple top-ranked wallets buy the same token within a time window. convergence_p0 (5+ wallets) is highest conviction; convergence_p1 (3-4 wallets) is moderate conviction.
Whale Move: A smart money signal fired when a single top-50 ranked wallet makes a large trade (default threshold: $50,000 USD).
Trader Score: A composite ranking metric (0-1) blending PnL efficiency (40%), Bayesian-adjusted win rate (30%), position consistency (20%), and activity (10%). Used to rank wallets in the leaderboard.
Smart Money Signal: A real-time notification generated when top-ranked wallets exhibit notable trading patterns (convergence buys, whale moves, sell signals). Delivered via polling or webhook.
Filter Terms
AI Labels: Pre-defined tags assigned by AI models to content, including:
- Topic Labels: Categories like arts_culture, business_entrepreneurs, science_technology, music, etc.
- Sentiment Labels: positive, neutral, negative
- Emotion Labels: joy, anger, anticipation, trust, etc.
- Moderation Labels: spam, hate, violence, harassment, etc.
- Web3 Topic Labels: web3_nft, web3_defi, web3_infra, etc.
Dynamic AI Labels: AI-determined labels based on user behavior and context. The system automatically selects appropriate labels for a user based on their interaction patterns.
Publication Type: The format of content, including image, audio, or video posts. Video posts can have additional filters for duration, orientation, and language.
Engagement Metrics: Measures of user interaction with content, including likes count, comments count, and shares count. Used to filter content by minimum or maximum engagement thresholds.
Geographic Location (geo_location): The geographic coordinates where content was created, specified as latitude,longitude pairs. Used for location-based filtering.
Author ID: The unique identifier (FID) of a content creator. Can be used to filter by specific authors, users followed by the viewer, or users followed by a specific user.
Technical Terms
Top K: The maximum number of items to return in a feed response. Typically ranges from 1 to 500, with a default of 25.
Return Metadata: A boolean flag that controls whether full post details (text, author, timestamps, engagement counts, AI labels) are included in the API response.
Source Feed: A field in the API response that indicates which feed configuration generated each item. Can include main feed identifiers, fallback feed identifiers, or promotion identifiers.
User-to-User Interaction: Actions between users, such as following, messaging, bookmarking, or blocking. These interactions help build the social graph and inform affinity-based recommendations.
User-to-Item Interaction: Actions between users and content, such as liking, commenting, sharing, bookmarking, upvoting, or downvoting. These interactions inform interest-based recommendations.
Interaction History: The record of all user interactions (both user-to-user and user-to-item) that is used to build user profiles and predict future engagement.
Following Graph: The network of users that a particular user follows. This social graph data powers affinity-based recommendations and social proof filtering.
Updated 20 days ago