aura-search is a high-performance, 100% offline search engine for visual media. it uses openvino™ to accelerate inference on local intel hardware (cpu/igpu/npu), allowing you to query your own videos and images using natural language.
| layer | technology | optimization |
|---|---|---|
| reasoning | llama-3.2-1b | openvino™ int4 (nncf) |
| vision | llava-v1.5-7b | openvino™ ir format |
| vector store | lancedb | apache arrow / lfa |
| embeddings | all-minilm-l6-v2 | hardware-mapped fp16 |
| processing | opencv | compute-efficient sampling |
the system is built to minimize memory overhead while maintaining sub-second retrieval speeds.
- ingestion: opencv extracts frames from local video files; llava generates semantic descriptions.
- indexing: descriptions are converted to embeddings and stored in lancedb, which leverages apache arrow for zero-copy memory access.
- querying: natural language queries are matched against the vector store; the results are synthesized by the local llama-3.2 model to provide context-aware answers.
- clone & setup:
git clone https://github.com/Udayscode/aura-search
cd aura-search
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
- run:
python3 main.py
- multimodal video/image support
- openvino™ hardware acceleration
- whisper-base integration for audio transcription
- layout-aware document (pdf/docx) parsing
this project was built to explore the efficiency of small language models (slms) on consumer-grade hardware. by using nncf int4 quantization, we reduced the memory footprint by ~70%, making complex multimodal search possible on standard laptops without external gpu dependencies.