Content Similarity Embeddings

Domain: Post targeting, AI distribution, user resonance detection Used By: Boost Targeting AI, SemanticTrustOracle, GeoOracle

The Content Similarity Embedding System forms the backbone of all AI-driven targeting, moderation guidance, and user-content matchmaking on the TRN platform. It encodes all content into vector form—text, audio, video, and image—so that semantic similarity, regional compatibility, and relevance scoring can be performed efficiently and accurately.

📚 Embedding Types

Each piece of content submitted to the platform is parsed and processed into embeddings using AI models. These embeddings are used by multiple systems in real time and are never stored in isolation—always connected to a CID and evaluated in secure off-chain environments.

Content Type
Embedding Process

Text

Full post text is tokenized, cleaned, and encoded

Audio

Transcribed and semantic features extracted

Video

Scene and language segments are separated and embedded

Image

Visual embeddings using standardized models

Mixed Media

Combined modality scoring and weighted vector fusion

🧠 Operational Flow

  1. User posts content

  2. The system generates embeddings for all applicable data types

  3. These embeddings are associated with the post CID

  4. Boost Targeting AI retrieves embeddings during campaign prep

  5. Embeddings are compared against user resonance vectors (view, bless, retrn history)

  6. Embeddings are also passed to SemanticTrustOracle and GeoOracle to test for violations or jurisdictional restrictions

🎯 Semantic Matching Logic

All targeting and moderation models use cosine similarity or equivalent distance measures to determine the resonance or dissonance between content and user/viewer profiles.

A threshold band defines what qualifies as a match:

  • Tight Match: 0.95+ cosine similarity (priority targeting)

  • Mid Match: 0.80–0.94 (balanced targeting)

  • Loose Match: <0.80 (fallback or extended reach)

Similarity ranges may be dynamically expanded or contracted during boosting depending on projected reach.

🧭 Targeting Use

Embeddings are central to:

  • AI targeting of boosted content

  • Determining if an ad or post matches a user's engagement history

  • Estimating campaign reach before launch

  • Updating user embeddings post-engagement (reinforcement learning)

🛡 Moderation Use

The embeddings are used to:

  • Trigger moderation flags when similarity matches banned content (via SemanticTrustOracle)

  • Enforce country-level content rules without revealing user-sensitive data

  • Provide automated appeals insights if content is falsely burned or blocked

🔁 Real-Time Feedback Loop

Every engagement updates the system’s understanding of content relationships:

  • Blessing a post increases the resonance signal for similar content

  • Burning a post lowers score for that vector across the user’s future feed

  • Retrning a post reinforces its position within that user’s interest profile

  • These signals also impact the LottoModule for quality validation

🔍 Embedding Refresh & Cache Policy

  • Embeddings are cached per post until deleted or expired

  • If post content is edited, new embeddings are generated and override old

  • Boosted posts reuse the original post’s embeddings unless altered

🧩 Connected Modules

  • Boost Targeting AI

  • SemanticTrustOracle

  • GeoOracle

  • LottoModule

  • ModerationLog

PreviousBoost Targeting AINextSemanticTrustOracle

Last updated