Curate data and monitor Insights

Curators turn App Logs or uploaded files into maintained Datasets. Workspace Insights summarizes how your Apps are being used and how their generations perform. Together, they help you move from observed production behavior to evaluation and training data.

Create a Curator

Open an App, select Curators, and select Create Curator. A Curator always writes to one target Dataset. You can name a new Dataset during setup or select an existing one.

Enter a Curator Name, choose the target Dataset, and set Data source to Logs or Files.

Curate Logs

For Logs, configure the collection before selecting Submit:

  • Run Hourly enables automatic hourly processing. When it is off, you can run the Curator manually.
  • Filter selects eligible App Logs. The form shows the current matching count.
  • Sample Rate processes between 1% and 100% of eligible Logs.
  • Split labels output as Training, Evaluation, Both, or Optimize (80% training/20% evaluation).
  • Limit stops automatic collection after the target Dataset reaches the specified item count.

An hourly Curator processes Logs added since its previous run. Use Run when you want to process the available set immediately.

Curate files

For Files, configure:

  • Chunk Size, from 10 to 1,000
  • Splitter, either Sentence or Markdown
  • Use text for, either Input or Output in each Dataset item

After creating the Curator, select Run, upload one or more .txt, .pdf, .csv, .md, .docx, .htm, or .html files, and confirm the run. Klu uploads the files, chunks their text, and writes the resulting items to the target Dataset.

Refine the curation pipeline

The same form provides optional transformations:

  • LLM Filtering uses your instructions to decide whether an eligible item enters the Dataset.
  • Remove Duplicates applies a three-level similarity threshold from Keep fewer through Balanced to Keep more.
  • Relabel Input with LLM and Relabel Output with LLM regenerate those fields. Instructions can reference {{input}} and {{output}}.
  • Generate more for each item creates from one to ten additional examples per accepted item using your generation instructions.

These transformations change the source material. Inspect the resulting Dataset before treating generated labels or examples as evaluation truth. Use a Dataset limit when an hourly Curator could otherwise grow the collection beyond the size you intend to review or train on.

Run and monitor a Curator

Open a Curator to see its target Dataset, current batch-job status, and run history. The Curators list also shows the last run and whether automatic hourly runs are enabled.

Select Run to process Logs or upload files. A successful submission reports Curator is running; processing continues as a batch job. If a run fails, Klu displays the returned error. Adjust the source, filters, or transformation instructions and run it again.

Select the target Dataset name to review the produced items. From there, you can correct examples, adjust splits and tags, create an Eval, or start a Fine-tune.

Read workspace Insights

Select Insights in the workspace navigation. Insights activates from platform use; workspaces with fewer than 29 generations show an activation message instead of charts.

Use the date-range control to choose the reporting period. Use Select an app to show All Apps or one App. The selection applies across the Insights page.

Insights includes:

  • Generations, with the previous period and percentage change
  • Helpfulness, based on generations classified as Helpful or High
  • Sentiment, including the share classified as Positive
  • Moderation, including generations recommended for review
  • Usage and Feedback over time
  • Top Topics, with an optional category filter
  • counts for Apps, Actions, A/B Experiments, and Fine-tunes
  • generation cost, generation count, feedback count, and errors
  • average input tokens, output tokens, and latency
  • top sources and models

Some cards depend on generated Insights or feedback. Empty cards or absent percentages mean Klu has no qualifying observations for that period and App selection. Treat these summaries as operational signals; inspect the underlying Logs before deciding why a metric changed.

Move from Insights to Evaluate and Fine-tune

Use Insights to choose where to investigate. For example, a rise in moderation recommendations or a weak helpfulness share can point you to an App and date range. Open that App's Logs, filter to the relevant generations, inspect their inputs, outputs, and feedback, then save reviewed examples as a Dataset.

Assign failure cases and corrected expected outputs to the Evaluation split, then create an Eval to measure candidate Action changes. Put approved training examples in the Training split before starting a Fine-tune. Keep held-out evaluation items out of training when you want a trustworthy post-training comparison.

Continue with Work with datasets, Evaluate prompt and model performance, and Fine-tune models.

API scope

Curator configuration, runs, and aggregate workspace Insights are product UI workflows in this guide. For programmatic Dataset and Experiment operations, use the public clients described in SDK exports. For generation ingestion and feedback, see API basics and API actions.