In-Browser & Qdrant AI Vector Search

Status: Coming Soon / In-Browser Active Development

We are actively developing client-side in-browser inference using Transformer.js (WebGPU) and optional self-hosted Qdrant vector backend integration.

Saved Posts Tracker is designing privacy-first deep learning models for visual understanding and facial identification without requiring expensive third-party AI APIs.

1. Multimodal CLIP Embeddings (In-Browser & Qdrant)

• Model: HuggingFace CLIP (Xenova/clip-vit-base-patch32) executed locally in-browser via @huggingface/transformers or on backend.
• Dimensionality: 512 floating-point vectors.
• Query Mechanism: Both image pixels and text descriptions are projected into the same latent embedding space.
• Vector Storage: Client-side vector index with optional Qdrant backend for large-scale self-hosted instances.

Example Natural Language Queries:

• "Moody neon cyberpunk street"
• "Warm wooden Scandinavian interior"
• "Minimalist typography posters with Swiss grid"

2. In-Browser Face Detection & Facial Descriptors

• Model: @vladmandic/face-api running on @tensorflow/tfjs in-browser backend.
• Dimensionality: 128-dimensional facial descriptor vectors.
• Clustering: Automatically detects faces in saved images, extracts biometric embeddings, and allows filtering all posts containing matching individuals with zero biometric data leakage.

Reindexing Vectors CLI

When vector indexing is activated, you will be able to run:

npm run reindex:vectors