6-Layer Observation

Every LLM call in the pipeline is captured across six observation layers. These layers split into two groups: three that stream in real time during execution, and three that are available for post-mortem inspection after the query completes.

Inner Layers (Trace Only)

These layers expose what happened inside the LLM call. They are available in the Trace tab after the pipeline finishes:

• Prompt – the full prompt sent to the model, with [System] / [Instructions] path-format separators
• Thinking – chain-of-thought reasoning, visible when thinking is enabled for that role
• Raw Response – the unprocessed model output before any parsing

Outer Layers (Live + Trace)

These layers stream in the Live tab as each stage completes, and are also captured in Trace:

• Parsed Output – structured fields extracted from the raw response (action, tool, verdict, scores, etc.)
• Execution – tool calls, search results, and intermediate state
• Observation – what the agent observed from tool execution (chunk counts, relevance scores, quality labels)

Why Six Layers

A single LLM call involves more than just input and output. The prompt is assembled from multiple sources (system instructions, retrieved chunks, conversation history). The model may reason through a chain-of-thought before responding. The raw response needs parsing into structured actions. Those actions trigger tool executions that produce observations fed back into the next step. Six layers capture this full lifecycle without collapsing distinct concerns into one view.

Per-Role Coverage

All four pipeline roles – ReAct, Grader, Judge, and Fallback – emit the same six layers. This means you can inspect the prompt, thinking, and raw response for a Grader scoring call with the same interface you use for a ReAct reasoning step or a Judge verdict.