Skip to content

Professional services

Need expert help? RAGcon offers:

  • RAGflow Support
  • RAGflow Hosting
  • Moodle Support
Talk to RAGcon →

AI provider (RAGflow)

Repository & issue tracker

Component: aiprovider_ragflow · Release: 0.5.0 · Requires: Moodle 5.0–5.2 · Depends on: — (root of the suite)

The RAGflow AI provider plugs into Moodle's core AI subsystem and connects Moodle's text AI actions to an external RAGflow instance. Answers are produced by a RAGflow chat assistant over the OpenAI-compatible endpoint, so they are retrieval-augmented and grounded in the assistant's knowledge base rather than a plain LLM. It is also the shared backend of the whole suite: it hosts the chat engine, session/long-term memory, source citations and a secure download proxy that the Tutor, Search and Helpdesk plugins consume. Install and configure it first.

Features

  • Core AI actions: serves generate_text, summarise_text, explain_text, each answered by the configured RAGflow assistant.
  • Assistant-driven model: the assistant's own model and knowledge base are used; a live dropdown lists the tenant's assistants and annotates each with its knowledge-base document count (or "no knowledge base — LLM proxy only"), so you can pick a RAG assistant or use RAGflow as a plain LLM.
  • Knowledge-base scoping & metadata filtering: answers can be filtered to this Moodle (course + site), the whole KB, or external/shared documents, and restricted to the current course or the user's enrolled courses via a document metadata field.
  • Source citations: optionally returns the source documents behind an answer, built from the model's own [ID] citations (only the documents actually used). They are numbered per answer as [answer.source] (e.g. [1.1], [2.1]), shown on a Sources: line at the end of the answer and as a linked list — linking to the Moodle activity when known, otherwise through a secure proxy.
  • Secure download proxy (download.php): streams a RAGflow document server-side so the API key never reaches the browser; per-click signed, time-limited links (token mode) or token-less context-authorised links, with a strict content-type allowlist (nosniff, forced attachment for anything but PDF/PNG/JPEG/GIF/WebP/plain-text).
  • Conversation (session) memory: for the Helpdesk drawer, RAGflow keeps the conversation so follow-ups have context and the transcript is restored on return.
  • Long-term memory: optional per-user durable facts via RAGflow's native Memory API (opt-in, off by default; disabled in private/incognito mode; cleared with the user's data).
  • Answers in the user's Moodle language (profile or course-forced language).
  • Per-user rate limiting (20 requests/minute) and a robust completion path (120 s timeout, one retry on network/5xx, not on 4xx).
  • Diagnosable failures: the real technical cause (HTTP status, RAGflow {code, message}, embedding/context-window errors) is surfaced to permitted users and to the dashboard; a coarse error type feeds usage analytics.
  • Usage events (metrics only, no content): chat_completed / chat_failed / search_performed for the optional dashboard.
  • Scheduled task: Prune stale RAGflow conversation sessions (daily) removes sessions unused past the retention period (default 30 days).

Configuration

The provider is configured as an AI-provider instance (Moodle's AI subsystem), plus per-action config. There is no classic settings page.

Provider instance — Site administration → AI → AI providers → RAGflow API provider

Setting Type Default Meaning
RAGflow API key (apikey) password (required) — Your RAGflow API key (RAGflow → User settings → API). Sent as the Bearer token and used to list assistants.
RAGflow base URL (baseurl) URL (required) — Base URL of your RAGflow instance, e.g. https://ragflow.example.com.
Download link lifetime (seconds) (tokenttl) integer 60 (min 15) How long a signed source/file download link stays valid. Links are minted on click, so a short lifetime is safe.

The provider counts as configured only when both API key and base URL are set.

Per-action config — per Generate / Summarise / Explain text action

Setting Type Default Meaning
RAGflow chat assistant (chatid) select (live) / text (required) — The assistant to answer with. Its model and knowledge base(s) are used. Pick a KB assistant for RAG, or a KB-less one to use RAGflow as a plain LLM.
System instruction (systeminstruction) textarea action default Instructions prepended as a system message to steer the response.
Document source (datasource) select thismoodle wholekb (no filter), thismoodle (filter: course + site), or external (shared documents only).
Restrict to course(s) (coursescope) select off Current course or The user's enrolled courses. Hidden for whole KB / external.
Course metadata field (coursemetadatafield) text course_id RAGflow document metadata field holding the Moodle course id.
Include sources (includesources) checkbox off Return source documents and append them as a linked list.
Extra parameters (JSON) (extraparams) textarea empty Optional JSON merged into the request body (e.g. {"extra_body": {"reference": true}}). Validated as JSON.

Where the chat/drawer settings live

The Tutor block and Helpdesk placement each carry their own copy of the chat settings (assistant, greeting, memory, sources, …) — configured on those plugins, not here — while the provider hosts the shared engine that reads them. See the Tutor and Helpdesk pages.

Capabilities

Capability Default roles Purpose
aiprovider/ragflow:viewerrordetails Manager, Teacher (site admins always) See the technical cause of a failed chat (a Details disclosure). The cause can reveal server-side internals (e.g. a RAGflow embedding error or HTTP 502), so it is withheld from ordinary users — enforced server-side, RISK_CONFIG.

Privacy

By default the plugin stores no personal data in Moodle. Prompts (and, with long-term memory on, the ongoing conversation and remembered facts) are sent to and stored in RAGflow (a third-party processor). A small local table references RAGflow conversation sessions for the Helpdesk memory. On user deletion, the plugin deletes those sessions and forgets the user's RAGflow memory. Users can self-disable long-term memory via private/incognito mode in the chat.

See also: Moodle version specifics for public/ layout and version-guarded error handling.

Note: This documentation was created with the help of AI. Spotted an error, an omission, or something unclear? Please report it in the ragcon-docs GitHub repository so we can fix it.