Skip to content

Sherpa Advisor

Sherpa Advisor is the context-aware scientific assistant in SpectraSherpa Cloud. It is aware of the active project, workflow graph, node settings, plots, metrics, reports, and generated artifacts, so exchanges can be specific to the analysis in front of you rather than generic chemometrics tutoring.

Use Advisor as a scientific review partner: ask it to explain the math behind a method, interpret model statistics, compare preprocessing choices, inspect what a plot is saying, or suggest alternatives when a result looks weak. When Gen Mode is enabled and you explicitly ask for a new approach, Advisor can propose an alternative workflow sheet while preserving the original analysis.

For a first use

Open the project, workflow, result, or report you want to discuss before asking your question. Name the response you need: an explanation, a check, a comparison, or a draft. If you have not opened the relevant evidence, assume Advisor cannot interpret it reliably.

A practical conversation pattern

For senior users, the most useful prompt supplies the scientific decision and the evidence boundary:

  1. State the question. For example: “Is the residual pattern consistent with a baseline problem, a sample-group effect, or an extrapolation issue?”
  2. Name the result. Say which workflow, run, target, response units and validation population are in scope.
  3. Ask for alternatives and checks. Request two or three plausible explanations, the evidence that would distinguish them, and the next check to perform in the Workbench.
  4. Verify the answer. Compare every numerical or scientific claim with the plot, table, saved parameters, provenance and validation design.

Good prompts are specific and evidence-bound:

  • “Explain why the held-out RMSE differs from calibration RMSE. Identify the population and denominator used by each.”
  • “Compare SNV and MSC for this dataset. List the assumptions, what each may hide, and a diagnostic plot I should inspect before choosing.”
  • “Draft a results paragraph that separates measured values, interpretation, and unresolved limitations. Do not invent units or validation claims.”

Junior-user note

Start with “What does this plot show?”, “What is this node doing?”, or “Which fields should I check before I run this?” Then ask Advisor to name the relevant samples, units, population and validation split. Do not ask “Is my model good?” without specifying what good means and which samples support the claim.

Use It For

  • choosing a first workflow
  • explaining PCA, PLS, SIMCA, MCR, validation metrics, and model diagnostics
  • interpreting scores, loadings, residuals, confusion matrices, calibration plots, report sections, and saved model artifacts
  • comparing preprocessing options against the current data context
  • drafting labeled narrative text for reports
  • asking what to inspect next
  • translating UI concepts into spectroscopy language
  • generating alternative workflow proposals when Gen Mode is enabled

Drafts and workflow proposals

Advisor-generated narrative is a draft. Edit it against the saved result before including it in a report. When Gen Mode is enabled, a proposed workflow is a new reviewable sheet; inspect its nodes, parameters, data source, target and validation design before running it. The original sheet is not your backup for an unreviewed proposal—save or export important work first.

Advisor does not make a proposed workflow scientifically valid merely because it can be executed. You remain responsible for checking leakage, group separation, units, sample identity, preprocessing fit scope, model applicability, and external or held-out claim boundaries.

Keep In Mind

Advisor is an aid. It can suggest useful checks, but it can also be incomplete or wrong. Advisor explanations, narrative text, and workflow proposals should be treated as labeled scientific assistance, not as validated results. Confirm conclusions with spectra, plots, metrics, sample IDs, file provenance, validation design, lab records, and domain knowledge.

Advisor behavior depends on the hosted service configuration and may be rate limited in demo mode. Hosted Pro uses deployment-managed providers and does not accept user-supplied LLM keys. If another enterprise deployment exposes a separately qualified BYOK or customer-selected provider, model behavior, availability, data-handling terms, and cost controls depend on that supplier too.

Do not paste secrets, credentials, access tokens, or information that your deployment policy does not permit into chat. Confirm the deployment's retention and sharing policy before discussing regulated or confidential material.