Skip to content

Professional services

Need expert help? RAGcon GmbH offers:

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

RAGflow Search

RAGflow file search — find documents by meaning, no hallucinations; understands scans, images and tables

Repository & issue tracker

Component: block_ragflowsearch
Display name: RAGflow file search
Requires: Moodle 5.0–5.2
Depends on: aiprovider_ragflow

More than a keyword search

This block searches the content of your files — including scanned PDFs, images and tables — by meaning, not just their titles. See Document understanding.

A Moodle block that adds a search box to any page and searches one or more RAGflow knowledge bases semantically and lists the matching source documents. It is retrieval only: no answer is generated by an LLM, so results cannot be invented and always link to real documents. The basic search is fast and cheap; an optional rerank model re-orders the results for better precision, at a little extra latency. The search itself runs in the shared AI provider, and each block chooses its knowledge base and scope.

RAGflow file search block on a Moodle course page with an empty search field The Search block adds a semantic search box to any page.

Easy setup

  1. Install the RAGflow AI provider first
  2. Add the Search block to a page
  3. Select the knowledge base(s)
  4. Tune the result quality

Features

Search results with relevance scores and text snippets linking to documents Results carry a relevance score and a snippet and link to the source document.

Search block showing the not configured hint Users who may configure the block see a hint instead of a dead search box.

  • Placeable on any page (courses, the Dashboard, the front page, …).
  • Semantic search over one or more RAGflow datasets, chosen per block instance.
  • One ranked list of matches (documents and media together, ordered by match score). Each row shows a file-type icon, the file name and the match score; the matching text excerpt appears on hover (info icon). Each result links to the source document through the provider's secure download proxy (the RAGflow API key never reaches the browser).
  • Quality filters for better matches: a relevance floor, a relevance cliff and a result cap remove weak hits; images and media keep a lower floor so they are not lost; a semantic weight helps full-sentence questions match by meaning; and an optional rerank model improves precision further (see Result quality).
  • Scope: whole KB or current course (course scope filters by a configurable metadata field; on pages without a course the whole KB is searched).
  • Login-gated (no content for logged-out users or guests); one instance per page.
  • Helpful empty states: a not configured hint for users who can add the block, a no knowledge bases available message if the provider isn't enabled, and a knowledge base was deleted in RAGflow notice when a search returns nothing because a configured dataset no longer exists (privileged users additionally see which dataset).

Use cases

Three ways teams use the Search block — expand an example. Pair any of them with the RAGflow Moodle Connector to keep the searchable knowledge base in step with the documents maintained in Moodle, with no manual re-upload.

Enterprise — find the right document fast

A company keeps policies, product docs and process guides in Moodle, and staff waste time hunting for the right file.

The Search block finds it by meaning — a query like "parental leave" surfaces the actual policy PDF and the passage that matched, as a link, with no generated answer to second-guess. Scanned forms and posters stay findable because RAGflow reads their content (OCR and a vision model), not just their filename.

Universities — course library & past papers

Students search across course handouts, lab manuals and past exam papers by meaning rather than exact title — useful when the file they need is a scanned handout or an oddly named PDF.

Every result links straight to the source document, so a student verifies the original instead of trusting a summary.

Public sector — auditable information retrieval

A public body offers a knowledge base of regulations, forms and procedures. Retrieval-only search returns the authoritative document itself — no LLM step, nothing that can hallucinate — which is exactly the property a compliance or citizen-information context needs.

It runs on a RAGflow instance under the body's control — self-run or privately hosted — so the corpus stays there.

Configuration

Almost all configuration is per block instance, and only site administrators may set it (non-admins see an "only site administrators can choose the knowledge base" message, and their saves cannot clear the choice). The only site-wide setting is an optional logging toggle (below).

Per-instance block config (Configure this RAGflow file search block)

Configuration form of the RAGflow search block showing datasets, scope and thresholds All settings are per block instance: datasets, scope, rerank model and thresholds.

Setting Type Default Meaning
Knowledge base(s) (config_datasets) autocomplete (multiple) none (required) The RAGflow dataset(s) this block searches. Select one or more; the block does not search until at least one is chosen.
Search scope (config_scope) select Whole knowledge base Whole knowledge base or Current course only (matched via the course metadata field). On pages without a course, the whole KB is searched.
Course metadata field (config_coursefield) text course_id RAGflow document metadata field holding the Moodle course id. Only used when scope is Current course only.
Rerank model (config_rerankmodel) select None (off) Optional. A dropdown of the rerank models available in your RAGflow (fetched live; a hint is shown if none is configured). When set, RAGflow reorders the candidates with a cross-encoder for markedly better precision. Choose None for plain vector/keyword ranking.
Semantic weight (config_vectorweight) number 0.7 Balances meaning-based (vector) vs. literal keyword matching (0–1). Higher = questions asked in full-sentence form match by meaning; lower = literal keyword matching dominates. RAGflow's own default (0.3) is keyword-heavy and scores sentence questions poorly.
Minimum relevance (config_minsimilarity) number 0.35 Text results below this score (0–1) are dropped. Higher = fewer/stricter, lower = more/looser. Images/media keep their own, lower floor.
Maximum results (config_maxresults) number 5 Upper bound on the number of text (non-media) results kept. Images/media are kept under their own lower relevance floor (up to 3 more) and shown inline in the single match-ranked list.
Relevance cliff (config_cliffratio) number 0.6 Keep a result only while its score stays within this fraction (0–1) of the top hit. Lower = keep more mid-ranked results; 0 = off.
Setting Type Default Meaning
Write log data (logtomoodle) checkbox off Write a concise usage/error entry (metrics only) to the Moodle standard log per search; independent of the RAGflow dashboard.

Result quality (fewer, better matches)

Open rerank model dropdown listing the models available in RAGflow The rerank dropdown lists the rerank models available in your RAGflow.

The search is tuned to return a short, relevant list instead of a fixed number of hits. The defaults work well; the semantic weight, minimum relevance, maximum results and relevance cliff can be tuned per block (see the config table above), and reranking is opt-in.

  • Relevance floor: matches below a minimum score are dropped, so weakly related documents no longer clutter the list.
  • Relevance cliff + result cap: once the scores fall away from the best hit (or the cap of 5 is reached), the list stops. A query with two strong matches returns two, not a padded eight.
  • Semantic weight (sentence questions): the hybrid search blends meaning-based (vector) and keyword scoring. A higher semantic weight (default 0.7) lets meaning dominate, so a question asked as a full sentence matches the right document instead of only its literal keywords. RAGflow's own default is keyword-heavy, which scores sentence questions poorly; raising the weight is the main fix.
  • Images and media appear in the same list: images and other media embed with low text similarity, so a single floor would wrongly hide them. They keep a lower floor so they still appear; all results (documents and media) then appear in one list ordered by match score, so media sit with their (lower) scores rather than being dropped or diluting the ranking.
  • Optional reranking: set a rerank model (above) and RAGflow reorders the retrieved candidates with a cross-encoder. This is the single biggest precision gain: the truly relevant documents rise to the top and the weaker ones fall below the cliff. It is opt-in (needs a rerank model in your RAGflow) and adds a little latency.

In short: shorter, more trustworthy result lists; no irrelevant filler; images stay findable; and, with reranking on, noticeably sharper ordering. Sensible defaults apply from the start.

Capabilities

Capability Default roles Purpose
block/ragflowsearch:addinstance Teacher, Manager Add a search block to a page
block/ragflowsearch:myaddinstance Authenticated user Add the block to the Dashboard / My home

Note

Choosing the knowledge base and scope is restricted to site administrators (not a capability), because a dataset is a site-wide resource.

Roles & permissions (who can do what)

The search block has no dedicated "use" capability: anyone who can see the block can search it and open results. Permissions only govern who may add the block and who sets up its knowledge base.

Role What this role can do
Site administrator Add the block anywhere and choose its knowledge base and scope (a site-wide dataset resource, admin-only, not a capability). Also sets the site-wide block setting.
Manager · Teacher (editing) Add a search block to a course page (:addinstance) and set its display options; cannot choose the knowledge base / scope (site-admin only).
Any authenticated user Add the block to their own Dashboard / My home (:myaddinstance).
Non-editing teacher · Student · anyone who can view the block Search and open results; no extra permission is needed beyond seeing the block.
Guest / not logged in Can search only where guest viewing of the page is allowed; otherwise no access.

Privacy

The block stores no personal data. Search queries are sent to the configured RAGflow service by the AI provider to retrieve matching documents; source links go through the provider's secure proxy.

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.