Trellis

Generates abstractive summaries of selected text passages or full documents using language models, allowing users to specify summary length and detail level. The system processes highlighted or full-text content through an LLM pipeline that extracts key concepts and synthesizes them into coherent summaries without requiring manual note-taking or external tools.

Converts selected or full-document text to audio using text-to-speech synthesis with adjustable playback speeds (typically 0.5x to 2x), allowing asynchronous consumption of reading material during commuting, exercise, or multitasking. The system likely uses cloud-based TTS APIs (Google Cloud TTS, Azure Speech Services, or similar) with client-side playback controls and speed normalization.

Provides an integrated annotation system allowing users to highlight text, add notes, and tag passages with metadata (e.g., 'key concept', 'question', 'definition') without fragmenting the reading experience. Annotations are stored in a structured format (likely JSON or database records) linked to document position and content, enabling retrieval, filtering, and export workflows.

Automatically generates targeted discussion questions and comprehension prompts based on document content using prompt engineering or fine-tuned LLMs. The system analyzes text structure, key concepts, and learning objectives to create questions at varying difficulty levels (recall, comprehension, analysis, synthesis) that guide deeper engagement with material.

Provides a unified reading environment that layers AI capabilities (summarization, TTS, annotation, questions) directly into the document view without requiring external tools or context switching. The interface likely uses a web-based document renderer (possibly PDF.js or similar) with embedded UI controls for each AI feature, maintaining reading state and document position across tool invocations.

Accepts multiple document formats (PDF, DOCX, EPUB, web URLs, plain text) and normalizes them into a unified internal representation suitable for AI processing and rendering. The system likely uses format-specific parsers (PDFKit or similar for PDFs, pandoc-like converters for DOCX) and OCR for scanned documents, extracting text and metadata while preserving document structure.

Maintains reading state (current page/position, scroll location, time spent) across sessions and devices, allowing users to resume reading without manual bookmarking. The system likely stores reading progress in a user database with timestamps and device identifiers, enabling cross-device synchronization and reading history analytics.

Enables full-text and semantic search across a user's library of documents and annotations, using keyword matching and embedding-based similarity search to find relevant passages. The system likely indexes documents and annotations using vector embeddings (from models like OpenAI's text-embedding-3 or similar) stored in a vector database, enabling queries like 'find all passages about machine learning ethics' across multiple documents.