Skip to content

Multi-Stage LLM Workflows

This advanced feature guide explains how to combine processors. For a practical starting task, see Dictate Prompts for AI Tools, Capture and Organise Voice Notes, or Hear Summaries and Responses.

Goal

Use a profile's processor graph to run several focused transformations in sequence or send the same text to multiple processors and endpoints.

This can support cleanup, extraction, classification, cataloguing, and final synthesis without forcing one model prompt to do every job.

How the Graph Works

Each profile stage selects one input:

  • Original text, meaning the transcript after word replacements.
  • The output of an earlier stage.

The profile separately selects one Final Output. That selected text continues into Auto Tag, trigger routing, paste, saved outputs, Notes, or an HTTP response.

When Stages Start

When a stage's selected input becomes available, the stage becomes ready. If several stages directly select that same input, OpenScreech initiates their requests in profile order. It initiates the next request as soon as the preceding request is submitted, without waiting for that request's response or result, so the sibling requests continue concurrently.

Each stage output becoming available is a separate readiness event. Downstream groups associated with different stage outputs initiate independently and can overlap; they do not pass through one graph-wide submission queue. Result ordering exists only where the profile explicitly makes one stage use another stage's output.

Pattern 1: Sequential Refinement

Original → Cleanup → Structure → Final Output

Use this when each task depends on the previous result. For example:

  1. Cleanup corrects punctuation without changing meaning.
  2. Structure converts the cleaned text into headings and lists.
  3. Structure is selected as the final output.

This is easier to test than one large prompt because each processor has a narrow responsibility.

Pattern 2: Parallel Specialists

                  ┌→ Extract Actions
Original → Cleanup├→ Classify Topic
                  └→ Create Summary

Set the three specialist stages to use Cleanup as their input. They can begin together after cleanup finishes and can use different local or remote endpoints.

Possible specialist roles include:

  • Extract action items and owners.
  • Classify the text for cataloguing.
  • Produce a short summary.
  • Convert the text into a task or issue template.
  • Check terminology or policy requirements.

Pattern 3: Required Side Checks

                  ┌→ Requirements Formatter → Final Output
Original → Cleanup┤
                  └→ Risk Reviewer

Use a side branch when its result supports an independent consumer such as Screechback or when you want additional processing that must not delay the primary transcript. If a check must gate delivery, place it in the selected final-output ancestry.

A stage accepts only one input source, so a later stage cannot directly combine two separate branch outputs. If the final text must contain several analyses, use one processor that produces that combined result or pass the required material through a single sequential chain.

Keep this limitation in mind when planning "swarm" behavior. OpenScreech can fan work out to multiple endpoint agents, but it is a text-processing graph rather than a general multi-agent message bus.

Important Output and Failure Behavior

  • Every enabled stage executes, including branches that are not selected as final output.
  • A failed stage blocks only its descendants; successful sibling branches continue.
  • A failed stage prevents primary delivery only when the selected final output depends on it.
  • Only the selected final output enters the normal downstream workflow.
  • Individual side-branch results are not currently written as separate artifacts.
  • A disabled stage makes no request and passes its selected input through, preserving downstream connections.
  • Stages that become ready from the same input initiate in profile order, then continue concurrently without waiting for sibling results.
  • Stages becoming ready from different inputs initiate independently and can overlap.

These rules make side branches appropriate for independent background work. Put required transformations and gates on the selected final-output chain.

Example: Catalogue a Spoken Research Note

Suggested processors:

Processor Input Purpose
Faithful Cleanup Original Correct punctuation while preserving facts
Research Summary Faithful Cleanup Produce a concise summary
Topic Classifier Faithful Cleanup Identify a controlled topic label

If the summary is selected as final output, the classifier still runs but its result is not separately saved. To retain classification in the finished text, use a final processor that produces both the summary and classification in one structured result.

Example: Prepare a Coding Request

Original → Remove Filler → Requirements Formatter → Final Output
                         └→ Risk Reviewer

The risk reviewer can validate the same formatted instruction against a separate endpoint without delaying or invalidating an independent successful final-output branch. Its result is not independently persisted, so give it an explicit consumer such as Screechback or treat it as optional background processing.

Building Safely

  1. Create and test each processor independently in the LLM Post-Processing Library.
  2. Add one stage to the profile and verify its output.
  3. Add dependencies one at a time.
  4. Confirm the profile's selected Final Output.
  5. Put required transformations on the selected final-output chain; use side branches only when their independent processing or consumer is useful.
  6. Duplicate shared processors before changing them for one workflow only.
  7. Review which endpoints receive transcript or selected-text content.

Quick Tips

  • Parallel branches isolate unrelated work as well as improving specialization. A failed side branch blocks only its descendants and does not invalidate an independent successful final-output branch.
  • If an analysis must survive, include it in the selected final text or a downstream combined result. Side-branch text is not saved as an independent artifact.

See Also