RAG Knowledge Base

Description

Retrieval-Augmented Generation (RAG) system with semantic search using vector embeddings. DOCUMENTS: upload and index PDF, DOCX/DOC, PPTX (with slide text, speaker notes, and optional visual analysis of slides with little text), Excel (XLSX/XLS), CSV, JSON, HTML/HTM, TXT, and Markdown. PDFs that are tabulated price catalogs are automatically detected and converted into a queryable SQLite table (tested on catalogs of hundreds of pages and thousands of product references). MULTIMEDIA: upload images (PNG/JPG/GIF/WEBP), audio (MP3/WAV/OGG/M4A/AAC/FLAC), and video (MP4/WEBM/MOV). A multimodal model catalogs each file (description + tags) and indexes it for semantic search. The gallery shows video posters (generated with ffmpeg), playback with seeking (HTTP Range), and the chat embeds playable images, audio, and video from the retrieved sources. DATABASES: upload SQLite files (.db/.sqlite/.sqlite3) or SQL dumps (.sql, which are imported into a queryable database). The query engine is read-only (SELECT/WITH only). In chat, the AI automatically generates SQL from your questions using the schema and a data sample from your databases (supports multiple databases at once). ORGANIZATION: group knowledge into collections with subfolders up to 3 levels deep. A reusable tree selector lets you filter by collection; selecting a folder automatically includes its subfolders, and disabling a folder disables its entire subtree for search. COLLABORATION (optional): with KB Workspaces enabled, each user can create named workspaces and share specific collections from their KB with other people by email (invitation with token + acceptance). Three roles: VIEWER (queries and uses shared collections in chat/RAG), EDITOR (additionally uploads, edits, and deletes documents, which remain under the KB owner's identity), and AUDITOR (additionally sees the workspace's search history, attributed to who searched, and never exposes the owner's private searches). A member can leave a workspace at any time. A shared workspace can also be used as a KB source for an Agent or an OmniChat bot, queried with the owner's identity and collection scope. Double-switch governance and legacy-safe: the superadmin enables the feature at the platform level and, on top of that, each owner opts in for their KB. With the feature disabled (default), each KB is 100% personal and all related functionality is hidden and inactive. INTEGRATION: the AI automatically uses your knowledge in chat (with a toggle), and it is also available for Agents and OmniChat with configurable privacy scope. It also includes deterministic data analysis (exact calculations over Excel/CSV/JSON/SQLite, without an LLM) and usage statistics.