Rebuild the vector index from all markdown files in the knowledge base. Scans companies/, people.md, projects.md, meetings/ and indexes all content into ChromaDB with rich metadata for semantic search and filtering. Call this after adding or updating knowledge base files.

Incrementally update the vector index — only re-index changed files. Compares file MD5 hashes to detect additions, modifications, and deletions. Much faster than index_knowledge() when only a few files changed.

Search the knowledge base using semantic similarity with optional filters.

Args:
    query: Natural language search query (e.g. "金融AI", "RPA项目进展")
    type: Filter by entity type: person/project/company/meeting (optional)
    company: Filter by company name (optional, partial match)
    person: Filter by person name (optional, partial match)
    top_k: Number of results to return (default 5)

Generate an image from a text description using Gemini (nano banana). Use this when the user needs illustrations, diagrams, icons, or any visual content. The image will be saved to the current working directory.

Args:
    prompt: Description of the image to generate
    filename: Output filename (default: output.png)

Generate a complete PPT presentation from a topic and optional context. Plans an outline using Gemini, generates slide images, and outputs a PDF.

Use this when the user wants to create a PPT/presentation. The /ppt skill
calls this tool automatically with knowledge base context.

Args:
    topic: The presentation topic (e.g. "润信科技 介绍", "AI在金融领域的应用")
    context_text: Optional reference content to inform the slides (max 3000 chars used)
    page_count: Number of slides to generate (default 5)

Transcribe an audio/video file to text using local Whisper. Automatically detects multiple speakers using pyannote and matches against registered voiceprints. Unknown speakers are labeled as "未知说话人1", "未知说话人2", etc.

Use this when the user provides an audio/video file and wants:
- Speech-to-text transcription
- Meeting minutes / meeting notes
- Interview transcription
- Any audio content converted to text

Args:
    file_path: Absolute path to the audio/video file
    model_size: Whisper model size (default: large-v3 for best Chinese support)

Register a person's voiceprint from an audio file. The audio should contain only one person speaking clearly for 20-30 seconds. The voiceprint will be saved and used to identify this person in future meeting transcriptions.

Args:
    name: The person's name (will be used as label in transcriptions)
    file_path: Absolute path to an audio file with the person's voice