LLM Enrichment

Puts a language model in the pipeline and keeps what it answers.

The LLM node sends asset metadata to a Large Language Model and stores the model’s response as enrichment — for example a classification, a set of suggested tags, or a description. Multiple prompts can be configured, and each prompt adds its own output port, so downstream nodes connect to the one answer they care about rather than to the node as a whole.

Kind

llm

Applies to

Any (operates on the asset’s filename / metadata)

Input ports

media — the asset, whose filename and stored fields the prompts reason over

Output ports

result_{promptId} — one port per configured prompt, so a summary prompt produces a result_summary port you can connect to Text-to-Speech or Sentiment Analysis

Requirements

An LLM provider service reachable from the worker. The node supports an OpenAI-compatible (vLLM) endpoint; CPU on the worker, the model runs in the service.

Persists to

asset_json_comp per prompt + asset_node_result ledger

Configuration

The LLM (Large Language Model) settings panel: OpenAI URL

Figure 1. The node’s settings in the pipeline editor

Set these in the panel above, or in the node’s options block in a pipeline definition:

Option Meaning

Option	Meaning
`openaiUrl`	Base URL of the OpenAI-compatible LLM server (llama.cpp, vLLM, Ollama `/v1`, …)
`prompts`	A map of prompt definitions. Each one adds a `result_{promptId}` output port, which appears on the node in the editor as soon as the prompt is defined

openaiUrl

Base URL of the OpenAI-compatible LLM server (llama.cpp, vLLM, Ollama /v1, …)

prompts

A map of prompt definitions. Each one adds a result_{promptId} output port, which appears on the node in the editor as soon as the prompt is defined

Seeing it run

Turn on Debug Mode and every node keeps what it produced, on the card itself. Below is a real run of this node over pexels-jack-sparrow-5977265.mp4.

The strip on the card lists what each output port carried — result_summary.

Use Cases

Auto-classification — categorise assets from their metadata or transcript.
Tag suggestions — propose tags for review.
Description generation — draft human-readable summaries stored alongside the asset.

Looking for something else?

All documentation Feature overview