Find prefabs by meaning, not filename. Fully local Unity Sentis embeddings, SQLite index, instant SIMD search. No API keys.AI Semantic Search is an Editor-only tool and is 100% render-pipeline agnostic: it works the same in Built-in, URP and HDRP projects. All AI inference runs locally via Unity Sentis (CPU or GPU compute backend). Requires com.unity.ai.inference (Sentis) 2.3.0+, resolved automatically by UPM.AI Semantic Search lets you find project assets by describing them in plain language - "heavy axe", "small green plant", "loud explosion" - instead of remembering exact filenames.Everything runs locally in the Editor: no API keys, no internet, no cloud.Features:Natural-language search - BERT-style embeddings (MiniLM-L6-v2) computed locally with Unity Sentis.Prefab intelligence - understands names, components (Lights, VFX, Physics), folder/category context, materials and mesh geometry.SIMD-accelerated cosine similarity + text-match bonus, with a sensitivity slider for fine control.SQLite storage - project-local index (Library/), gitignored, rebuildable at any time.Extensible - add custom metadata extractors and asset kinds via a small API (Metadata), no core changes.Included sample - MiniLM ONNX model, BERT tokenizer config and medieval demo prefabs so the tool works immediately after import.How it works:Import the sample, then open Window > SnivelerCode > Semantic Search.Assign the model and tokenizer once (Embedding tab).Press Check and Index in the Prefabs tab.Search in plain English.Requirements:Unity 2022.3 LTS or newer (tested on Unity 6 / 6000.5).com.unity.ai.inference (Sentis) 2.3.0+, resolved automatically by UPM.Compute shader support for the GPU backend (CPU backend available).Architecture:Editor-only UPM package (sniveler-code.dev.semanticsearch): UI Toolkit window, custom DI container (MiniContainer).Embeddings: Unity Sentis ONNX runtime (com.unity.ai.inference 2.3.0+), MiniLM-L6-v2 (float32, ~90 MB) - CPU or GPU (compute shader) backend.Search: SIMD-accelerated cosine similarity with text-match bonus and per-project sensitivity setting; branch-free vector math (Unity.Mathematics)Storage: SQLite (sqlite-net) at Library/SnivelerCode_SemanticIndex.db - asset index, properties, metadata; content-hash-based atomic category re-bake (FNV-1a).Categories: 253 built-in keyword databases (General 77, Detailed 138, Components 27, Materials 11); project-editable JSON with automatic re-bake.Extensibility:Metadata API - add custom metadata extractors and asset kinds without core changes.Category databases support an optional project-local examples field (typical asset names) that enriches category embeddings.Testing & quality:30 EditMode unit tests (vector math, tokenizer, SQLite storage, property roundtrip, category databases).CI: GitHub Actions test matrix on Unity 2022.3 LTS and Unity 6 (6000.x).Dependencies:com.unity.ai.inference (Sentis) 2.3.0+ - Unity package, resolved automatically. No Asset Store dependencies.(Note: the MiniLM embedding model is a third-party open-source model (Apache-2.0), redistributed with the samples; it is not an AI-generated artifact.)Store Description: AI language models assisted in drafting, refining, and formatting the marketing copy, including the package summary and feature descriptions for the Asset Store page.Documentation: AI was used to help structure, proofread, and write the technical documentation (README and PDF guides) to ensure the instructions are clear, professional, and easy for users to understand.(Note: The core C# codebase, ECS architecture, and DOTS integration were developed manually without AI generation).
