AIBOX-K3 AI Computer Seewise Application Usage

799959745

Seewise

Seewise is an intelligent video search engine that supports both local video uploads and RTSP camera streams. It automatically analyzes video content and enables users to quickly locate specific clips through natural-language search.

Key Features

• Intelligent Video Understanding: Automatically analyzes video frames and generates natural-language descriptions for each frame.
• Multiple Search Modes: Supports semantic search, keyword search, and hybrid search.
• Flexible Video Sources: Works with both uploaded local videos and live RTSP streams.
• Real-Time Processing Feedback: Displays processing progress directly in the user interface.
• Bilingual Support: Provides both Chinese and English interfaces, with multilingual search prompts.
  

Platform Support

Platform & OS	Supported
K1 Buildroot	❌ No
K1 OpenHarmony	❌ No
K1 Bianbu LXQT/GNOME	❌ No
K3 Buildroot	❌ No
K3 OpenHarmony	❌ No
K3 Bianbu LXQT/GNOME	✅ Yes

Technical Architecture

Seewise uses a client-server architecture with three layers:

• Frontend Layer: A modern web interface for video upload, search, and real-time processing updates.
• Application Service Layer: Handles business logic for video processing, search, and data management.
• AI Model Layer: Performs image analysis, text vectorization, and search result reranking.
  

System Architecture Diagram

Installation and Deployment

Install the Debian Package

For production deployments, installing the local package with apt is recommended:

1. sudo apt update
   
2. sudo apt install seewise-2

Copy the code

If dependency issues occur, run the following command:

1. sudo apt-get install -f

Copy the code

After installation, seewise-2.service is created and enabled automatically.

• Default root directory: ~/.seewise-2
• Runtime dependencies include llama.cpp-tools-spacemit, spacemit-onnxruntime, and python3-spacemit-ort
  

Quick Start

• Web Access: After installation, open http://<board-ip>:8084 in a browser.
• Desktop Access: Open the application menu in the lower-left corner, search for Seewise, and click the icon to open the corresponding web page.
  

Model Download and Parameter ConfigurationDownload Models

For production use, download models from the Settings page in the Seewise web UI.

Default model directories:

• VLM: ~/.seewise-2/models/vlm/fastvlm-mm-0.5b-q4_1/
• Embedding: ~/.seewise-2/models/embedding/
• Rerank: ~/.seewise-2/models/rerank/
  

It is recommended to use the default recommended models.

Configure Parameters

The main configurable settings include frame extraction and retrieval parameters:

• Frame Extraction Settings: Define the frame sampling interval, which determines how frequently frames are extracted from the video.
• Image Size Settings: Define the model input image size, which affects both retained detail and inference time. Larger images preserve more detail but require more processing time.
• Retrieval Settings: Control the search strategy, including whether reranking is enabled.
  

Video Upload and RTSP Retrieval Workflow
Local Video Upload Workflow

• Select Upload Video in the frontend interface.
• The backend saves the file to data/videos/.
• FFmpeg extracts frames and generates output in data/frames/ and data/thumbnails/.
• The VLM generates frame-level semantic descriptions.
• The embedding service generates vectors and writes them to the vector index.
• The processed video is indexed and ready for search.
  

RTSP Live Stream Workflow

• Enter the RTSP address in the frontend interface and start the connection.
• The backend uses RTSPCapture to record the stream and save it as an MP4 file.
• FFmpeg then extracts frames and generates thumbnails.
• Processing progress is pushed to the frontend in real time through WebSocket.
• When processing is complete, frame data is written to both the database and the vector index.
  

Search Workflow

• Upload a video or connect to an RTSP stream, then wait for processing to complete.
• During processing, a progress indicator appears in the lower-right corner of the video player, and keyframe previews appear below the search box.
• Enter a natural-language query in the search box and press Enter to search for relevant content.

  For best results, use the retrieval configuration options available on the Settings page.

  
  

Note: Due to current environment limitations, searching is not supported while video processing is still in progress.

The application includes two built-in demo videos and their corresponding keyframe semantics for quickly demonstrating scenarios such as relevance-based search.

Logs and Cache Management
Log Locations

• Backend log: ~/.seewise-2/logs/backend.log
• Frontend log: ~/.seewise-2/logs/frontend.log
• Model logs: ~/.seewise-2/logs/vlm.log, ~/.seewise-2/logs/embedding.log, ~/.seewise-2/logs/rerank.log
• systemd log streaming: journalctl -u seewise-2.service -f
  

Cache and Data Directories

• Video files: ~/.seewise-2/data/videos/
• Extracted frames: ~/.seewise-2/data/frames/
• Thumbnails: ~/.seewise-2/data/thumbnails/
• SQLite database: ~/.seewise-2/data/video_search.db
• Vector index: ~/.seewise-2/data/vector_store.db
  

To clear cached data, follow these steps:

• Stop the service:
• 1. sudo systemctl stop seewise-2.service

  Copy the code
  
  
• Back up or remove the cache directories and files
• 1. rm -rf ~/.seewise-2/data/videos/*
     
  2. rm -rf ~/.seewise-2/data/frames/*
     
  3. rm -rf ~/.seewise-2/data/thumbnails/*
     
  4. rm -f ~/.seewise-2/data/video_search.db
     
  5. rm -f ~/.seewise-2/data/vector_store.db

  Copy the code
  
  
• To download the models again, delete the model directory and repeat the model download process.
  

FAQ
Q: Why does the service fail to start after installation?

A: First check systemctl status seewise-2.service. If the issue is dependency-related, confirm that llama.cpp-tools-spacemit, spacemit-onnxruntime, and python3-spacemit-ort are installed.

Q: Why can't the models be downloaded, or why is the model service not starting?

A: Check the following logs to identify the cause:

• ~/.seewise-2/logs/vlm_8071.log
• ~/.seewise-2/logs/embedding_8072.log
• ~/.seewise-2/logs/rerank_8073.log
  

Q: Why does the RTSP connection fail?

A: Verify the RTSP URL, network connectivity, and camera status. If RTSP recording fails, review the backend log and confirm that the required port is not already in use.

Q: Why does video upload fail, or why is processing slow?

A: The maximum file size is 500 MB. Large videos may require more processing time. Also verify that sufficient disk space is available in data/videos/ and data/frames/.

Q: Why are the search results inaccurate?

A: Try switching the retrieval mode between semantic, keyword, and hybrid search, and enable reranking if needed. If the issue appears to be model-related, re-download the models and restart the model services.

Q: How do I clear the cache and start over?

A: Stop the service first, then delete ~/.seewise-2/data/video_search.db, ~/.seewise-2/data/vector_store.db, ~/.seewise-2/data/frames/, ~/.seewise-2/data/thumbnails/, and ~/.seewise-2/data/videos/.

Q: A popup says search failed during querying. What causes this?

A: Check ~/.seewise-2/logs/backend_8084.log for a vector-dimension mismatch error, such as:

1. File "app/routes/search.py", line 49, in search
   
2. File "app/services/search_service.py", line 272, in hybrid_search
   
3. File "app/services/search_service.py", line 43, in semantic_search
   
4. File "app/services/vector_store.py", line 119, in query
   
5. ValueError: shapes (1024,) and (768,) not aligned: 1024 (dim 0) != 768 (dim 0)

Copy the code

If this happens, restart the application and switch back to the correct embedding model.