Features

A self-hosted RAG pipeline for WordPress — ingestion, embeddings, vector search, chat, providers, MCP, a conflict-free widget, and a full admin UI.

📚 Knowledge base ingestion

  • Import URLs — paste one or many (one per line, or title,url CSV). Pages are fetched, boilerplate is stripped (<nav>, <footer>, <script> …), the readable text is extracted, then chunked, embedded and indexed.
  • Import files — PDF (via smalot/pdfparser), DOCX (native ZipArchive + word/document.xml parser), TXT, Markdown. Upload through the WordPress media library or paste a URL.
  • Index your own WordPress content — pick which post types (posts, pages, custom) become searchable. Auto re-indexes on save_post, removes on wp_trash_post. The permalink is stored as the citation source.

🧠 Embeddings & vector search

  • Configurable chunk size, overlap, and minimum length (sentence-aware splitter that preserves formatting — no sanitize_text_field mangling).
  • Embedding providers: OpenAI, OpenAI-compatible (Together / Azure / vLLM / LM Studio), Cloudflare Workers AI (bge-*, bge-m3), Ollama (local, nomic-embed-text).
  • Vector stores:
    • MySQL — uses native VECTOR(n) columns + DISTANCE(..., 'COSINE') when MySQL ≥ 9 is detected; transparently falls back to JSON storage + PHP cosine similarity otherwise.
    • Cloudflare Vectorize — full REST v2 client (insert, query, delete-by-ids, create-index). A “Create index” button auto-detects the dimension from your embedding model.

💬 Chatbot

  • SSE streaming (fetch + ReadableStream) for live, token-by-token answers, with a non-streaming /chat/sync fallback.
  • Citations — built from retrieval metadata and appended as a Sources list. Per-answer toggle in Chat settings.
  • Reasoning / extended thinking — per provider (reasoning_effort on OpenAI, thinking.budget_tokens on Anthropic, <think> blocks on DeepSeek/Ollama models). Shown in a collapsible UI section.
  • Tool calling — when an LLM emits tool calls the chatbot executes them in a bounded loop.
  • Rate limiting — per-IP request throttling via transients.
  • Conversation history — server-side + localStorage session continuity.

🤖 LLM providers

OpenAI · OpenAI-compatible · Anthropic Claude · Cloudflare Workers AI · Groq · Ollama (local). Per-provider API key, base URL, model. “Fetch models” and “Test connection” buttons in the admin. See Providers for setup details.

🔌 MCP integration (Model Context Protocol — client mode)

  • Connect to external MCP servers (streamable HTTP or SSE transport) from the MCP tab.
  • Discover & cache each server’s tool list.
  • Discovered tools are offered to the LLM via function calling; the chatbot runs a bounded tool-call loop (max 5 iterations), invoking tools/call on the MCP server and feeding results back into the conversation.
  • Tool invocations surface as a 🔧 Using tool: X badge in the chat stream.

See the MCP guide for details.

🎨 Frontend widget

  • Conflict-proof — every element is scoped inside #openrag-widget, every class is prefixed openrag-, all theming happens through --openrag-* custom properties on the root. No leaked styles, no global selectors.
  • Two delivery modes: floating widget (auto-injected, toggleable) and [openrag_chat] shortcode for inline embedding.
  • 4 preset themes — Light, Dark, Ocean, Sunset — plus per-color overrides.
  • Logo, bot avatar, bot name, welcome message, launcher position — all configurable.
  • Vanilla JS, mobile-responsive, markdown rendering with HTML escaping (no XSS).

🛠 Admin UI

  • Dashboard — totals (documents, chunks, chats, 👍/👎), system status (active provider + vector store mode), recent activity.
  • Knowledge Base — Documents/Files, URLs, WordPress Content tabs. Add forms, status badges, per-item reindex/delete, chunk-preview modal, bulk CSV URL import.
  • Chats — paginated list of every conversation with search, date filter, feedback icons, a detail modal (message, reply, reasoning, citations) and CSV export.
  • Settings — tabbed: General · LLM Providers · Embeddings · Vector Database · Indexing · Chat · Appearance · MCP. Nonces + sanitization throughout.

⚙️ Background processing

  • Action Scheduler (the library WooCommerce ships) — one job per document/post, retry and exponential backoff handled automatically.
  • On-request mode alternative — process immediately via AJAX, ideal for small imports.