This service is behind a shared key. It is exchanged once for a session
cookie, so the key itself never lives in this page.
Docgen Agent
Drives a RAGFlow document-generation agent one turn at a time, and renders what it
returns — charts included.
Projects
One named project holds two RAGFlow datasets (created automatically as
"<name> - spec" and "<name> - evidence"), a workflow, a requirements document and
one document run — everything below drives the same /api/projects/*
endpoints an API caller would use, end to end, without opening RAGFlow. "Generate a document"
further down is the lower-level, manual version of the same machinery, still useful for
driving datasets you created yourself.
Datasets
Workflow
Picking one installs it immediately.
Spec documents
Client context, a source map, canonical facts — anything the writer should read as
specification rather than evidence. Goes into the same dataset as the workflow and the
generated requirements document.
Evidence
Requirements
Describe the document you want; one sentence is enough. Built, checked, and stored
straight into this project's spec dataset — it is not shown here at all; see the
traceability matrix below once a run starts.
Run
"Decline & finish" stops the run as it stands -- no more sections will be written -- and
marks it done, so Export below works immediately on whatever exists. It costs no turn: there
is nothing to ask the agent, only a decision to record.
Traceability matrix
Checks
Each finding is either MEASURED (arithmetic, certain) or the canvas's own JUDGED check
(a model's opinion). Choose Fix it for what should change,
Leave it for what should not, or Decide later to come
back to it. Every accepted fix in a section becomes ONE revise turn, not one per finding.
Includes the figure-source, references (or sources-consulted) and traceability
appendices by default — assembled, never generated, so nothing in an appendix can be
invented. "Export document" downloads it straight away; "Save to exports/" writes the
same file to this server's exports/ folder instead, for opening or sharing later.
Document
Read-only here -- the server is the source of truth for every stored section. Use
Checks and Revise on the left to change content; nothing typed here is saved.
Advanced: manual mode (classic, one dataset/agent at a time)
Generate a document
The agent writes from its own two RAGFlow datasets — one holding the
specification, one holding the evidence. The Python chart source it emits is rendered here
into images embedded in the HTML below. Needs RAGFLOW_DOCGEN_AGENT_ID and
RAGFLOW_AGENT_TOKEN in the shared .env.
Agent
Both boxes are per-request and are never saved, so switching between agents needs no
.env edit or container restart. A token typed here passes through the browser and this
server, so keep shared or long-lived tokens in .env instead.
Datasets
Created and filled from here — nobody needs to open RAGFlow. Tick which dataset holds
the spec (requirements and workflow) and which holds the
evidence; those ids are sent with every turn. Needs an agent whose
Retrieval nodes read them from the request — one with datasets pinned in the RAGFlow UI
instead ignores the ids, so ticking nothing is also fine there.
Not loaded. Press Refresh.
Workflow library
The method the document is written against. Installing one puts it straight into the ticked
spec dataset — the file is read on the server and handed to RAGFlow, so nothing passes
through this page. Installing again replaces the previous one rather than adding a second,
because two workflows in one dataset means the writer follows two incompatible methods and
nothing reports it.
Picking one installs it immediately into the ticked spec dataset.
Build a requirements document
Only if you do not already have one. Describe the document you want — what it is, who it
is for, what it must cover — and the agent writes a requirements document in the same
format a hand-written one uses, checks it, and puts it straight into the ticked spec
dataset. One sentence is enough. It is not shown here and not downloadable; it stays on the
server and goes only into your dataset. Needs an agent with the requirements branch.
"Save" writes what you see, figures included, and only this service renders those —
RAGFlow's own Export gets the markdown, so charts arrive there as Python source.
Open a saved .html from the link above rather than off disk: it styles itself from
this server.
Editable. A new section is APPENDED here; a revision replaces the last one. Both
rebuild the document from the sections received, so hand edits are best made once
writing is finished — "Save editor contents" writes exactly what you see.