AI Semantic Search for Unity

AI Semantic Search is a high-performance Unity Editor extension that finds your project assets by meaning instead of exact filenames. Describe what you need — "heavy axe", "small green plant", "loud explosion" — and get the matching prefabs back.

Everything runs fully locally: embeddings are computed in the Editor with Unity Sentis and the index is stored in SQLite. No API keys, no internet, no telemetry.

Features

• 100% local AI — runs on your machine with Unity Sentis and the bundled MiniLM model (~24 MB, uint8-quantized).
• Prefab search by meaning — understands names, components (Lights, VFX, Physics), folder/category context, materials and mesh geometry.
• Fast vector search — SIMD-accelerated cosine similarity with a text-match bonus for instant results.
• Adjustable precision — Sensitivity slider from recall-first (default 25) to strict matching (60–70).
• Extensible — add new asset kinds or custom metadata via the Metadata<T> base class.

Quick Start

1. Install the package: Window → Package Manager → + → Add package from tarball… and select sniveler-code.dev.semanticsearch-1.1.0.tgz from the downloaded zip.
2. Import the sample: Package Manager → AI Semantic Search → Samples → "Base AI Model & Demo Assets" → Import (bundled MiniLM model, BERT tokenizer and medieval demo prefabs).
3. Open the tool: Window → SnivelerCode → Semantic Search.
4. In the Embedding tab assign Model = MiniLM_uint8.sentis and Vocab = tokenizer.json. Leave Tokens = 128 and Backend = GPU (CPU works as a fallback).
5. In the Prefabs tab press Check then Index.
6. Search! Describe what you need in the Search tab.

Requirements

• Unity 2022.3 LTS or newer (developed and tested on Unity 6000.5.2f1)
• Unity Sentis (com.unity.ai.inference 2.3.0+) — installed automatically with the package
• GPU with compute shader support recommended (CPU backend available as fallback)
• Editor-only package — Windows / macOS editors, no build-target requirements

Notes

• The index lives in Library/SnivelerCode_SemanticIndex.db (project-local). If Library/ is cleaned, just run Check → Index again.
• Only BERT-like models (input_ids / attention_mask) are supported — .sentis and imported .onnx.
• Software license: MIT. The bundled MiniLM model is a uint8-quantized derivative of all-MiniLM-L6-v2 (Apache-2.0) — see THIRD_PARTY_NOTICES.md.

Documentation

sniveler-code.github.io/docs_semantic-search.html — full setup guide, configuration reference, extensibility and model notes.

Questions or feedback? Drop a comment or contact the author. Happy searching!
— SnivelerCode