Yuanzhu Technology

Semantic & multimodal search

Free information from its format, so every kind of content is easier to find.

Start from a natural-language description and find the information you really need.

Much important information exists as images, scanned files, recordings, videos and unstructured documents. Traditional search relies on file names, keywords and manual tags, and struggles to understand what is in a picture, what is said in a recording or how different media relate. Finding a video clip, an image with a particular object or a key remark in a recording usually means filtering and checking by hand again and again.

Yuanzhu uses Yuan MetaCore-K3 for semantic understanding and reasoning, combined with the knowledge retrieval and long-document processing of YottaMind-3.2, to build an information discovery flow with natural language as the entry point.

For images, audio and video, the system works with content parsing, speech transcription, visual feature extraction and indexing components to turn assets into searchable information. Users describe what they are looking for in natural language, and the search system matches assets using text semantics, visual features, transcripts and time information, returning the original locations for checking.

Users no longer need to remember exact file names or look through every asset. Through one natural-language search entry point they find document content and related images more easily and, where audio and video are indexed, jump to the matching clip, cutting repetitive cross-media searching.

One scenario, end to end

  • Bring in multimodal contentImport authorized text, image, audio and video assets into one content management entry point.
  • Parse contentUse content parsing components to extract document text, recognize image content, transcribe speech and record the time position of video clips.
  • Understand and retrieveYuan MetaCore-K3 helps understand the search intent, combined with the knowledge retrieval of YottaMind-3.2 and multimodal index components to match relevant content.
  • Locate and verifyReturn matching assets, key clips and their original locations, so users can open the content and check it.