Skip to main content
Pinecone Nexus records a trace for every query. The trace shows each step the model ran, the retrieval code it generated, and the tokens and time each step took, so you can understand a result or debug an unexpected one.

Open a trace

After a query returns, click Show trace. To trace an earlier query, open its session from the Sessions list and click Show trace on that query. A trace opens with the query text and the model that answered, and it ends with the final answer the caller received.

Summary bar

The summary bar at the top of the trace totals the run:
  • Steps: The number of reasoning steps the model took, plus the final answer.
  • Tool calls: The number of calls to the retrieval SDK.
  • Tokens: Input and output tokens for the whole query.
  • Latency: The time it took to produce the answer.
To see where the tokens went, click Token detail.

Token detail

  • Where the tokens went: How the tokens divide between the reasoning steps and the final synthesis, plus the total tokens billed.
  • New input by step: The new input tokens each step adds. Each step resends the whole conversation, so only the new tokens are new work, and the prompt cache can serve the rest.
  • Output tokens by step: The output tokens each step and the final answer produced.
  • Caching savings: How much input the prompt cache served versus how much was billed as new input. The more context a run reuses, the more of its input the cache serves.

Spans and inspector

The trace lists each span with its duration, from the model’s reasoning steps to the tool calls they made. The spans follow the agent loop described in Queries. Select a span to show its details in the inspector:
  • Reasoning step: Input and output tokens, timing split into decide time (the model reasoning) and execute time (the code running), and the code the model generated and ran.
  • Tool call: Its arguments (such as the SQL a query_db() call ran), its result (such as the number of rows returned), and its timing.