Platform GuideLLM Integration
LLM Integration Overview
The model registry and generation logs for the models your agents and judge run on.
Pipelines connects to large language models through a shared model registry — the platform-provided catalog plus any models you bring with your own API keys. These are the models your agents run on, that back the LLM judge, and that power any LLM-driven feature on the platform. Every call is recorded in detail for cost and latency observability.
In this section
Model registry
Platform-provided models plus bring-your-own-key providers — Fireworks, Together AI, Bedrock, HuggingFace, and any OpenAI-compatible endpoint.
Generation logs
Every LLM call records model, token counts, cost, latency, the tool-call trace, and the full prompt/response.
Key behavior
- Bring your own key. Add models from Fireworks, Together AI, Bedrock, HuggingFace, or any OpenAI-compatible API alongside the platform catalog. See Model registry.
- Tool calling. If the chosen model supports tools and your organization has at least one active Tool Endpoint, the model can call MCP or HTTP tools during generation in a multi-round loop. See Tools.
- Full observability. Per-call details — model, tokens, cost, latency, tool-call trace, and the full prompt/response — appear in the Data Explorer task detail, with pipeline-wide aggregates in the LLM Analytics tab. See Generation logs.
- Capability is not filtered. The model picker lists every available model; if the one you choose doesn't support a needed capability (e.g. tool calling), the related controls are disabled.