LLM Integration Overview

Pipelines integrates large language models (LLMs) directly into workflows. Most value-producing fields in Subtask and Review nodes can be configured to have their value generated by an LLM, with support for custom prompts, tool calling, and bring-your-own-key model credentials.

Capabilities

• LLM-generated fields — fields whose value is produced by an LLM at a context-appropriate time (at task creation, when the node activates, or on contributor demand) instead of by a human.
• Tool calling — if the chosen model supports tools and your organization has at least one active Tool Endpoint, the LLM can call MCP or HTTP tools during generation in a multi-round loop (see Tool calling).
• Prompt library — save, version, and reuse prompt templates across workflows, with pinning and A/B variants.
• Model registry — use platform-provided models or add your own via BYOK. Supported BYOK providers are Fireworks, Together AI, Bedrock, HuggingFace, and OpenAI-compatible (any OpenAI-style API endpoint).
• Generation logs — every LLM call records model, token counts, cost, latency, tool-call trace, and the full prompt/response. Per-task details appear in the Data Explorer task detail; pipeline-wide aggregates appear in Data Explorer's LLM Analytics tab.

Cross-cutting behavior

• Defaults cascade from the node. Model, temperature, and tool calling are set at the node level (in LLM Generation Settings on the Property Inspector) and can be overridden per field. System prompt is set per field (or at the shared-settings level of a batched LLM group).
• Model capability is not filtered. The model dropdown lists every available model; if the one you pick doesn't support a capability the field needs (e.g., tool calling), the related controls are disabled on that field.
• Review nodes still require a human verdict (Approve / Reject) even when every field is LLM-generated.
• Variables are chip-based. Custom and library prompts reference other fields via variable chips inserted from a picker (open with {{ or Insert Field); typing {{…}} by hand is sent as literal text. The picker only exposes non-hidden fields above the current one on the same node and non-hidden fields on upstream Subtask or Review nodes; upstream traversal skips fail edges and the reject side of review nodes. It never exposes task metadata or input.* / node.* / task.* namespaces.