`veralith.log()`

The explicit way to send a trace. It works with any stack or code shape, because you hand it the three values directly.

import veralith
 
trace_id = veralith.log(
    query="What is the Rule of 72?",
    context=chunks,
    response=answer,
    latency_ms=elapsed,   # optional
)

Signature

veralith.log(
    query: str,
    context: list[str] | list[dict] | list[ContextChunk],
    response: str,
    *,
    latency_ms: float | None = None,
) -> int  # trace_id

Parameter	Required	Description
`query`	yes	The user's question. Must be a non-empty string.
`context`	yes	The retrieved chunks. See Context formats for the three accepted shapes.
`response`	yes	The generated answer, as a string.
`latency_ms`	no	Your RAG pipeline's response time in milliseconds. Shown on the dashboard as RAG latency and feeds the p95 KPI.

Returns the trace_id immediately — the evaluation runs asynchronously.

Guarantees

log() follows the SDK contract: it's fail-safe (errors are caught and warned, never raised into your app), fire-and-forget (returns as soon as the trace is queued — no added latency), and a no-op if VERALITH_API_KEY isn't set.

Note: evaluation is hosted-mode only, so the sync=True option is not supported against the hosted backend. To grade a trace inline (no account, offline), use veralith.evaluate() instead.

Where to put it

Anywhere query, context, and response all exist in the same scope — the end of a RAG function, inside a web handler, after a stream finishes, in a notebook. If your code doesn't have such a single point, see Recipes by code shape.

Overview @veralith.trace