Return to Gemini Workspace Automation Lab
Gemini Workspace Automation Lab Help
Learning path Validate → repair → check evidence → approve → add to the simulated sheet
Start here
Learn how to check data extracted from a document before approving it for a spreadsheet. The opening response is prerecorded, so the exercise works without a Gemini key.
- Find the type errorChoose Invoice extraction and click Load invalid fixture, then Validate JSON. Expect
$.total must be a number.Find the quote marks around"113"in Structured JSON. - Repair and check the sourceChange
"total": "113"to"total": 113. Click Validate JSON. Then compare the total, currency, and evidence withTotal: 113.00 CADin the source. - Review and approveClick Open review. Enter
Workshop reviewerandChecked total and currency against the invoicein the reviewer and reason fields. Check I checked the source and supporting excerpt. Click Approve record, then Add to simulated sheet. Expect a row with113andCAD.
Detailed walkthrough, examples, and troubleshooting
Practice rejection and uncertainty
- Load the reference invoice again. Keep
totalnumeric but change it to100. Click Validate JSON. A numeric subtotal can pass the schema even though the source says the final total is 113. - Click Open review, enter a reviewer and the reason
Subtotal used instead of final total, and click Reject record. The simulated sheet gains no row. Inspect the decision in Review audit. - Try replacing
evidencewithPayment made yesterday. Validate again. The schema may pass, but the lab flags that the exact excerpt is absent.
Compare evidence and experiment with prompts
- Open Retrieval & evaluation. Search for
total CADand click Rank source chunks. Read the text offsets and matching words. - Click Compare fixtures and current response. Compare each row’s schema result, exact field matches, and excerpt presence. Read the document yourself before deciding whether an excerpt supports a claim.
- Return to Extraction lab. Change the extraction prompt to
Use the tax-inclusive total, not the subtotal.and click Save prompt version. In fixture mode, changing the prompt does not generate a new response. - Open Lessons & projects and search for Compare shape with evidence. Try the exercise with a second template such as Policy question answering.
Save, export, or try a live request
- Open Export & live setup → Save locally to keep the draft in this browser. The draft requires fresh review after reload. Download project JSON also includes your source text and audit, so review the file before sharing it.
- Choose
Code.gsorSETUP.mdunder Generated file, then click Download Apps Script ZIP to get the schema, review workflow, and instructions for an authorized Google test account. The ZIP excludes your pasted document. - Open Optional live Gemini request only if the site operator has configured it. Use fictional text, enter the supplied operator key, check the permission box, and click Request live response. Validate and review the returned candidate before adding any row.
If a button is unavailable
- Open review is disabled: repair any JSON errors and validate again.
- Approve record is disabled: open review and check that the evidence appears exactly in the current document. Enter a reviewer and reason and confirm the evidence check.
- Add to simulated sheet is disabled: approve first. Editing the prompt, document, or response sends the current record back to Draft.
What each area does
Extraction lab and workflow picker
Choose a fictional workflow, then load its invalid or reference fixture. A fixture is a prerecorded answer. The prompt can be edited and versioned without changing that answer. Start with the invoice to practice fixing one field.
Document and prompt
Source document is the evidence you will check. System instructions, Extraction prompt, and Example or constraints describe the requested work. Save prompt version stores the wording. Text and Markdown import are local. The optional live request is a separate action.
Response and review
Structured JSON is the candidate record. Validate JSON checks the format and field types. Open review lets you check source accuracy. Enter reviewer and reason, approve or reject, then use Add to simulated sheet to complete an approved record. Editing returns the candidate to Draft.
Simulated spreadsheet and audit
The sheet shows approved records after you explicitly add them. The audit lists validation and review decisions. Compare the two when testing a rejection: the audit records it and the sheet does not gain a row.
Retrieval and evaluation
Rank source chunks finds text with matching words. Read the offsets and verify which passage supports the claim. Compare fixtures and current response shows schema checks, exact field matches against the reference, and whether the evidence text appears. Read the source for factual accuracy.
Lessons and projects
Open a lesson to get a prediction and task. Try Compare shape with evidence after the invoice exercise. Templates switch the document, fields and recorded answer together. Use the approved policy exercise to practice distinguishing current and archived information.
Cost and AI coach
Cost uses hypothetical input and output prices per million tokens. Change one input at a time and compare estimates. The AI coach generates editable guidance to copy. Neither action contacts Gemini.
Export and live setup
Download Apps Script ZIP gives the reviewed Sheets workflow and setup files. Download project JSON includes your saved document and response, so check it before sharing. Live requests require a configured server and an explicit send action. Use the fictional sample for testing.
Button and field reference
Use these sections while you work through the quick start.
Extraction lab
- Workflow / Load invalid fixture / Load reference fixture
- Choose a fictional source and schema. Load the deliberately broken or worked prerecorded result. An edited prompt does not generate new fixture output.
- Source document / Import plain text or Markdown
- Read or replace the source used for evidence. Import stays local. The prerecorded results still belong to the original sample.
- System instructions / Extraction prompt / Example or constraints
- Define general behavior, the task and optional examples. These fields are used for a configured live request. In fixture mode you can practice editing them without sending anything.
- Save prompt version / Saved version / Restore version
- Keep prompt wording for comparison. Select an earlier version to restore it. Restoring the prompt requires fresh review.
- Structured JSON / Validate JSON / Validation rules
- Edit the candidate, check JSON and field types, and inspect the schema. Try changing total from the string "113" to the number 113.
- Apply reference repair
- Replaces your current candidate with the prerecorded correct example. Use this to compare your own repair.
- Open review / Reviewer / Reason / I checked the source and supporting excerpt
- Move a validated candidate to review, document the decision, and confirm the evidence check. A matching excerpt does not prove the claim is accurate.
- Approve record / Reject record / Add to simulated sheet
- Accept or decline the candidate. Only adding an approved record creates a row. Rejecting records a decision without adding one.
- Simulated spreadsheet / Review audit
- Inspect the resulting row and the sequence of decisions or validation failures. Changing the source, prompt, or candidate sends it back to Draft.
Retrieval and evaluation
- Query / Chunk size / Overlap / Rank source chunks
- Split the current text and rank excerpts by shared words. Use total CAD as the query, 180 as the chunk size, and 30 as overlap to start. The offsets identify where to find each excerpt.
- Compare fixtures and current response
- Separately compare schema compliance, exact reference fields, and excerpt presence. Check the source yourself to decide whether a claim follows.
Learning, costs, and exports
- Lessons & projects / Collection / Search
- Find a lesson or choose a workflow from Templates and Guided projects. Open a lesson to try its exercise and save a reflection.
- Requests / Input and output tokens / Input and output price / Practice quota / Calculate
- Estimate cost using editable assumptions. Prices are per million tokens. The quota comparison is fictional.
- AI coach / Goal / Generate prompt / Copy prompt
- Build editable troubleshooting or design guidance. It does not call a model or approve records.
- Generated file / Download Apps Script ZIP
- Inspect code, manifest, and setup. The ZIP includes the template schema and a review-gated Sheets script. It excludes your pasted document and credentials.
- Save locally / Download project JSON / Restore project
- Keep the draft and audit in this browser, export a copy including source text, or load a backup. A restored draft needs validation and review again.
- Optional live Gemini request / Operator access key / Send this source and prompt to the configured Gemini model / Request live response
- An authorized operator can submit a fictional document to a configured Gemini backend. The returned candidate still needs validation and review. Leave the panel closed for fixture-only practice.
- Switch theme
- Changes the lab appearance.