Where should a LlamaIndex workflow publish?

In the step that emits `StopEvent`. Publish the assembled document, then stop with the URL as the result, so callers get an address rather than a rendered document.

Why does this matter more for RAG output?

Because a RAG answer is only trustworthy with its sources attached. A rendered report can carry the retrieved passages inline with the claims they support; a chat message usually drops them.

Can an agent workflow publish mid-run?

Yes, and updating is the better pattern — publish once, then update the same report id as the workflow refines its answer, so the reader's link never goes stale.

Does this work with LlamaCloud or a deployed workflow?

Yes, and it is the only thing that does. A deployed workflow writes to storage you do not hand out; an HTTPS report URL is the artifact that escapes.

How do reviewers push back on a specific claim?

By commenting on the paragraph. Anchored comments put the objection next to the sentence and the citation it came from.

Share what a LlamaIndex workflow produced

A workflow retrieves, re-ranks, synthesizes and writes. The last step emits a StopEvent carrying a finished document — an evidence summary, a due-diligence memo, a literature review with citations.

handler resolves. The result is in memory. The analyst who commissioned it gets a paste into Slack that loses the formatting and every source link.

Publish in the stopping step

Workflows are event-driven and explicit, so publication is a normal step concern:

class ReportWorkflow(Workflow):
    @step
    async def finalize(self, ev: DraftEvent) -> StopEvent:
        html = render(ev.sections, ev.citations)
        r = httpx.post(
            "https://commareports.com/api/v1/reports",
            headers={"Authorization": f"Bearer {os.environ['COMMA_API_TOKEN']}"},
            json={"title": ev.title, "html": html},
        )
        r.raise_for_status()
        return StopEvent(result={"url": r.json()["url"], "id": r.json()["id"]})

The result is now two short strings. Anything downstream — a Slack notifier, a ticket comment, a calling workflow — handles an address instead of a document.

RAG output needs its sources visible

This is the part specific to LlamaIndex. A retrieval-backed answer is worth exactly as much as its citations, and citations survive badly in chat: links get flattened, node text gets truncated, and the reader is asked to trust a paragraph with a bracket number pointing nowhere.

A rendered page keeps them. Put each claim next to the passage it came from, keep the source metadata in the markup, and the reviewer can check the reasoning instead of accepting it. When they disagree, the comment lands on that paragraph rather than arriving as "the third point seems off" — see commenting on HTML.

Update, don't re-publish

Agent workflows iterate. If yours refines an answer over several passes, keep the report id from the first publish and update it:

httpx.patch(
    f"https://commareports.com/api/v1/reports/{report_id}",
    headers={"Authorization": f"Bearer {token}"},
    json={"html": improved_html},
)

One address, one comment history, and a link that keeps working after you shared it. The alternative is a reviewer commenting carefully on revision two while everyone else reads revision four.

Worth knowing

Deployed workflows have no reachable filesystem. Writing report.html in a LlamaCloud deployment or a container produces nothing — the same problem as publishing from CI.
Do not carry HTML through the event stream. Pass the URL between steps; the document does not need to travel.
reports:write to publish, comments:read to act on feedback — see scoped tokens for AI agents.
5 MB per report body.

Try it

Free — unlimited reports, commenters and revisions.

Read the API reference →