Product · Evaluation · observability

When an answer is wrong,
you can see exactly where

For a single question, you can see how many times the model was called, which documents were searched, which SQL ran, and where the time went. Quality scores build up automatically, so no one has to read every reply.

Request a demo → See integration specs

Every model call, search, and SQL run recorded TRACE

Time spent at each step LATENCY

Scores from automatic quality evaluation SCORE

Three points where you lose track of what went wrong

All you get is "the AI gave a weird answer"

Find the trace by conversation ID and confirm whether the cause was retrieval, a tool, or the model.

You can't tell whether switching models actually helped

Compare with automatic scoring, per-model trends, and Arena rankings.

LLM cost only shows up at the end of the month

Retracing a single run

Traces

Records the entire path behind a single answer — LLM calls, tools, retrieval, and guardrails.
Tree, timeline, and token views, with the slowest stage flagged automatically.
Click a node to see its input and output (prompt, retrieval results, SQL, tool response).
With a reasoning model, the reasoning summary and reasoning tokens are shown in the trace too.

Automatic quality evaluation

Turn on automatic quality evaluation for an agent and an evaluation model scores each answer from 0 to 1 on retrieval quality, faithfulness, and response quality.
User-uploaded attachments are excluded from the retrieval-quality score to keep scoring fair.
View score trends by model and by agent.

Feedback · Arena

Collects user feedback as structured, per-criterion data.
Ranks models with an Elo-based score from head-to-head answer comparisons.
Choose the Arena model and one of the registered models answers at random; which model it was is revealed only after the user rates the answer.

Usage

Aggregates tokens by model, user, group, organization, and agent.
View it as time series, treemaps, a day-of-week × hour heatmap, or a ranking table.
Register a per-model unit price (input, output, cache) to estimate cost, in your chosen display currency. Savings from prompt caching are also shown in the time series.
Look up conversation logs with filters.

Configuration health check

One admin screen checks whether LLM connections, embeddings, extraction engines, search, web search, code execution, Code Gateway, notifications, voice, images, encryption, and licensing are configured, and what each one affects.