aiPDF

Accepts PDF, EPUB, website URLs, and YouTube video links as input sources, routing each through a format-specific parser before initiating a background preprocessing pipeline. Users can begin querying documents immediately while preprocessing continues asynchronously, enabling non-blocking interaction. The system handles format detection, content extraction, and indexing in parallel without blocking the chat interface.

Implements a retrieval pipeline that matches user queries against document sections using relevance matching (likely semantic search via embeddings, though model unspecified), then passes matched sections to an LLM for response generation. Responses include 'detailed references' and are 'double-checked and backed by sources extracted from the uploaded documents,' enforcing grounding to document content only. The system prevents hallucination by constraining generation to information present in the source material.

Mentioned as a capability ('information extraction') but not detailed in documentation. Presumably, users can ask questions designed to extract specific information (e.g., 'list all dates mentioned in this document'), and the system returns structured or semi-structured answers. Implementation likely leverages the Q&A pipeline with prompt engineering to encourage structured output.

The product includes a charity donation feature where users can contribute to causes, with some portion of proceeds supporting charitable organizations. This is mentioned as part of the product's value proposition but implementation details (which charities, donation percentage, tax deductibility) are not disclosed. This is a business model feature rather than a technical capability.

Enables simultaneous conversation across multiple uploaded documents, allowing users to ask questions that synthesize information from different sources. The system maintains a 'multi-document chat' session (limited per tier: 1 free, 5 Dynamic, unlimited Flagship) and supports 'multi-document joins' (3 free, 5 Dynamic, 10 Flagship) where documents are queried together. Implementation likely extends the retrieval pipeline to search across multiple document indexes in parallel, then aggregate results before LLM generation.

Generates 'comprehensive' summaries that consider 'full context' of uploaded documents, likely using the same retrieval pipeline to identify key sections before LLM-based abstractive summarization. The system produces summaries grounded in document content rather than generic overviews, with implicit source tracking inherited from the Q&A capability.

Implements a multi-tier data retention policy where documents are automatically deleted after 1 month (Free), 6 months (Dynamic), or indefinitely (Flagship). Users can manually delete documents at any time. Storage is encrypted ('encrypted databases' mentioned, but vendor/location unknown). The system enforces tier-based retention as a hard constraint, with no option to override automatic deletion on lower tiers.

Provides Optical Character Recognition for image-based PDFs and scanned documents, with monthly page limits enforced per tier (50 pages Free, 500 pages Dynamic, 3000 pages Flagship). OCR is applied during preprocessing to extract text from image content, making it queryable via the Q&A pipeline. The metering suggests OCR is a resource-intensive operation with per-page costs.

Enforces monthly document upload limits per tier (2 uploads Free, 120 uploads Dynamic, unlimited Flagship), creating a paywall trigger at the second document for Free users. The system tracks upload count and resets monthly, preventing further uploads when quota is exhausted. This is a soft quota (users can upgrade) rather than hard technical limit.

Implements a monthly question limit per tier (550 questions Free, 5500 Dynamic, unlimited Flagship), tracking cumulative questions asked across all documents and sessions. When quota is exhausted, users cannot ask additional questions until monthly reset. This is a soft quota enforced at the API level, not a technical limitation.

Claims to 'support all languages' for document ingestion and querying, but implementation details are not disclosed. Presumably, the embedding model and LLM backend support multiple languages, but specific language coverage, script support (CJK, Arabic, Cyrillic), and accuracy across languages are unknown. This is a claimed capability without technical verification.

Provides a chat interface for interacting with uploaded documents, maintaining conversation history within a session. Each document or multi-document group has its own chat session, with session state managed server-side. The interface is standard conversational UI (similar to ChatGPT), but scoped to document context rather than general knowledge.