An AI zettelkasten that extracts ideas from articles, videos, and PDFs

Accepts articles (via URL or HTML), videos (via URL with transcript extraction), and PDFs as input sources, normalizing them into a unified text representation for downstream processing. The system likely uses content scrapers for web articles, video transcript APIs (YouTube, Vimeo), and PDF parsing libraries to extract text while preserving semantic structure, then standardizes output into a common format for idea extraction.

Uses an LLM (likely OpenAI GPT or similar) to analyze normalized content and extract discrete, atomic ideas formatted as individual zettelkasten notes. The system prompts the model to identify key concepts, claims, and insights, then structures them as standalone notes with clear relationships, enabling the core zettelkasten principle of linking ideas across sources. Implementation likely involves prompt engineering to enforce atomicity and semantic clarity.

Automatically identifies conceptual relationships between extracted ideas using embeddings or LLM reasoning, then generates bidirectional links between related notes. The system likely computes vector embeddings for each atomic note, performs similarity search to find related ideas, and optionally uses the LLM to validate or label relationship types (e.g., 'contradicts', 'extends', 'example of'). This enables the zettelkasten's core value: serendipitous discovery of connections across sources.

Stores extracted notes and relationships in a structured database or file system with full-text and metadata indexing, enabling efficient retrieval and browsing. Implementation likely uses a document database (MongoDB, SQLite with FTS extension) or file-based approach (Markdown files with YAML frontmatter) with indexed fields for source, date, tags, and relationships. This provides the foundation for querying and exploring the knowledge base.

Provides a user interface (likely web-based or CLI) to browse notes, search by keyword or metadata, and visualize relationships as a graph or outline. The system renders the zettelkasten as an interactive knowledge graph where users can click through related ideas, or as a hierarchical outline showing note connections. Implementation likely uses a graph visualization library (D3.js, Cytoscape, or similar) and a search interface with filters for source, date, and tags.

Supports importing multiple content sources (articles, videos, PDFs) in batch mode with asynchronous processing, queuing, and progress tracking. The system likely uses a task queue (Celery, RQ, or similar) to process imports in the background, preventing UI blocking and enabling efficient handling of large batches. Implementation includes job status tracking, error handling with retry logic, and optional webhooks for completion notifications.

Automatically preserves and indexes source metadata (URL, author, publication date, excerpt location) for each extracted idea, enabling citation generation and source verification. The system stores a reference to the original content for each note, allowing users to trace ideas back to their sources and generate citations in standard formats (APA, MLA, Chicago). Implementation includes metadata extraction during ingestion and citation formatting templates.

Supports multiple LLM providers (OpenAI, Anthropic, local Ollama, etc.) through a unified interface, allowing users to choose their preferred model or provider. Implementation likely uses an abstraction layer (e.g., LangChain, LiteLLM, or custom wrapper) that normalizes API calls across providers, enabling easy switching without code changes. Configuration is typically via environment variables or config files specifying provider, model, and API keys.