Digital Candidate / RAG LabBack to workspace
PRIVATE WORKSPACE · RESPONSE TUNING

Better answers. Visible evidence.

Compare retrieval settings, inspect the evidence, and measure each answer test.

Manual runs only

1. Open your candidate workspace

Opening a workspace or inspecting evidence can wake Azure SQL. This page never polls.

Ready when you are. No API requests have been made.

Model calls this session0
Reported tokens0
Estimated model cost · USD$0.000000
Retrieval engineSQL + keyword ranking

2. Choose a question and evidence scope

Uses committed SQL knowledge. Browser drafts and job-fit assessments are outside this lab.

A · CURRENT SERVER SETTINGS

Baseline

Uses the same selection settings as ordinary candidate Q&A.

B · YOUR EXPERIMENT

Tune one variable at a time

Presets save settings only, in this browser. Source and citation requirements remain enforced.

Model pricing and generation controls

Model deployment: not loaded

Enter your deployment's USD rates per million tokens. Leave unknown rates blank. Estimates exclude SQL, Search, hosting, storage, network and taxes.

Check Azure model pricing

Cached input is counted once. Reasoning tokens are already included in output tokens. Failed calls can incur charges even when usage is unavailable.

3. Inspect, then test an answer

Inspect makes no model call. Generate can make up to two model calls per profile, including one citation-recovery attempt. Each button runs only the selected profile.

A · Baseline result

Inspect the baseline to see which evidence reaches the model.

B · Experiment result

Change a setting, then inspect the experiment.

4. Review quality and keep the evidence

Mark relevant records in the results to compare retrieval. These are your judgments; unmarked records are treated as nonrelevant. A structural citation pass does not prove factual accuracy.

Reports include private evidence: download only when you want to retain it. Results otherwise stay in this tab. Exporting a preset does not change live answers.

What you are tuning

Candidate Q&A selects eligible SQL records, ranks term matches, builds bounded excerpts and sends a governed evidence packet to the model. This lab tunes that actual path. Azure AI Search serves application-audit questions separately; vector search, semantic reranking and index chunking are not active candidate controls.

Use the same question, audience and source versions for A/B comparisons. Repeat across several representative questions before promoting a setting. Open Azure Cost Management for account charges.