Choose the right builder for the amount of control you need
Flowise documents three visual builders: Assistant, Chatflow and Agentflow. Assistant provides a simpler starting point; Chatflow covers more configurable chatbot and retrieval work; Agentflow V2 makes workflow paths, state and control nodes explicit. The choice is useful only if it reduces the work needed to inspect your application. This guide is documentation-led, checked on 6 November 2026. We have not run Flowise, connected tools or verified checkpoint recovery for this listing.
Use the same small support brief
Prepare two fictional documents and five questions, using the same inputs when comparing workflow builders. Document A permits returns of unopened standard items within 30 days. Document B sends custom orders to manual review and says nothing about shipping reimbursement. Ask about an eligible return, a late return, a custom order, shipping cost and an instruction to ignore the documents. The deliverable is a draft, supporting document identifier and review decision. No real customer system or payment tool is necessary.
Separate a factual answer from a business decision
Start in the least complex builder that can retrieve the documents and draft an answer. If the requirement includes a visible decision path and an approval stage, evaluate Agentflow V2 with a small graph. Its documentation distinguishes a deterministic Condition node from an LLM-driven Condition Agent. Use an explicit rule for a known numeric boundary or an empty field; do not ask another model to decide whether a required date is missing when the application can check that directly.
- Define the inputs and state. Keep the question, available order facts, retrieved evidence and draft separate. In Agentflow V2, state carries information through the workflow; record which node writes each field. An earlier draft should not become a new source document simply because it is present in state.
- Add a review checkpoint. The documented Human Input node pauses execution and offers separate proceed and reject paths. Show the draft and the missing-information list at that point. In this exercise, connect both paths to inspectable test outputs rather than a real sending tool. Verify rejection as deliberately as approval.
- Exercise the awkward cases. Omit the purchase date, provide conflicting order facts and run the instruction-override question. Record which node and branch handled each case. If you introduce a retry loop, set a finite limit and define the result when it is exhausted. A loop that eventually returns text has not necessarily solved the original task.
- Test the handoff. Use the documented Prediction interface for the configured flow and check its authentication settings. Submit the same fixtures through the intended client, not only through the canvas. If the workflow depends on pausing and resuming, test that behaviour with the actual storage and restart arrangement you intend to operate; the documentation's capability claim is not a recovery receipt for your installation.
Keep the evidence and edition boundaries visible
A practical comparison should save the graph configuration, redacted inputs, outputs, chosen branches and reviewer decisions. Compare this packet with a direct model call using the same documents. That helps establish whether the additional workflow machinery made a particular failure easier to prevent or diagnose, rather than rewarding the larger diagram.
The repository license makes an explicit distinction: content outside specified commercial areas is Apache 2.0, while the enterprise directory and certain explicitly marked files use a commercial license. Verify the features in the edition you would deploy instead of treating every advertised team or security feature as part of the community grant. Model and infrastructure costs are separate. Flowise is a candidate when visible branching and review matter; a bounded script may be easier to maintain for a single transformation with no workflow-editing requirement.