Document understanding¶

The suite's answers and search are only as good as how well your documents are read. RAGflow does not just index file names and plain text — it parses the actual content of your files, including scanned pages, images and tables, and searches them by meaning. That is what sets it apart from Moodle's built-in search.

Not just text — the whole document¶

RAGflow's DeepDoc engine reads a document the way a person would:

Optical character recognition (OCR). Scanned PDFs and photographed pages become searchable text. Multilingual.
Layout understanding. It recognises the structure of each page — titles, paragraphs, tables, figures, headers/footers, references and equations — and reads it in the right order, so a heading stays with its section and page furniture (running headers/footers) is discarded.
Table structure. Tables are captured as rows and columns, not a flat blur of numbers.

Images and figures are understood, too¶

A vision model (configured as the optional img2txt model) describes the figures, diagrams and screenshots inside your documents. That description becomes searchable text, and the image itself is stored with the passage.
So an image can be found by search and shown as a source in a chat answer — not only the text around it.

Image understanding is optional

OCR, layout and table recognition are built in. Understanding the content of images additionally needs an img2txt (vision) model configured in RAGflow — see Set up RAGflow.

Choose how documents are parsed¶

DeepDoc (built in) — the default OCR + layout + table pipeline.
MinerU and Docling (RAGflow 0.25+) — alternative parsing methods for high-fidelity extraction, with options such as formula recognition and table recognition and multilingual OCR, selectable per knowledge base.
Content-aware chunking — templates for general text, papers, books, manuals, presentations, tables, Q&A, laws, e-mails and pictures split each document the way its type needs.

Meaning, not keywords¶

Retrieval embeds your query and your documents into the same vector space, so a question finds the passage that means the same thing even when it uses different words. An optional rerank model sharpens the order of the results.

How this differs from Moodle's search¶

	Moodle's built-in search	RAGflow (this suite)
Searches in	titles, text fields, metadata, forum posts	the content of the files themselves
Scanned PDFs / images	invisible	OCR + image description → searchable
Tables / diagrams	ignored	captured as structure / described by a vision model
Matches by	exact keyword	meaning (embeddings, optional rerank)
Returns	a link to the activity	the supporting passage + image, optionally an answer

Moodle's search is excellent at finding courses and activities by name. RAGflow is about finding what a document says — and answering from it.

What it means in practice¶

A learner's question is answered from the exact paragraph — even if it lives in a scanned handout or on a slide.
A diagram that explains a concept can surface in search, with the picture attached.
A trainer uploads the material as-is — no need to retype tables or transcribe scans.

Requirements¶

An embedding model is required (it builds and searches every knowledge base).
Image understanding needs an img2txt (vision) model; formula / high-fidelity extraction needs MinerU or Docling (RAGflow 0.25+).

See Set up RAGflow to configure the models, and the Search block and Tutor block to put it to use.

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.