docs/release_notes_2026_06.md

DemoPrep June 2026 Release Notes

This is the working prod-update summary for the recent DemoPrep changes. It is intended to be updated as fixes are tested and promoted.

Branch / Deploy State

• Current reconciled code line: develop / new-branch
• Test deploy target: hf-test/main
• GitHub source target: origin/develop
• Last known reconciled commit before the current local fixes: d58f32a

Data Quality And Dataset Generation

• Moved the app toward dataset-first generation:
  • research and scenario context produce a scenario contract
  • the dataset is generated coherently first
  • DDL is derived from the generated dataset
  • Snowflake loads the generated dataset directly
• Added scenario-family routing instead of one-off company-specific generators.
• Added and improved specialized or richer dataset paths for:
  • SaaS sales
  • retail sales
  • professional services
  • trucking/shipping/logistics
  • banking marketing and finance
  • restaurant/store operations
  • education enrollment
  • healthcare/life sciences
  • CPG/grocery finance and sales
• Improved data realism:
  • better dimension cardinality so datasets do not look toy-sized
  • coherent formulas and constraints for rates, funnels, margins, passengers, operations, and finance metrics
  • fewer random placeholder values and fewer impossible metric combinations
• Improved generated product, warehouse, and dimension names to reduce generic numbered labels such as Product 29 where possible.
• Added quality-run tooling and reporting:
  • e2e quality runs save JSON/Markdown artifacts
  • results can be written to ts_quality_results
  • the ThoughtSpot quality reporting liveboard reads from ts_quality_results rather than session_logs

Custom Context And Scenario Specificity

• Custom prompts and additional context are passed into scenario extraction so custom demos should preserve customer/domain nouns rather than falling back to generic matrix defaults.
• Scenario extraction is intended to drive public/custom demos; deterministic routing is a fallback and guardrail, not the main intelligence layer.
• Fixed cases where industry/use-case combinations routed to the wrong scenario family, including several logistics, finance, and CPG paths.

ThoughtSpot Deployment Reliability

• Chunked ThoughtSpot table imports to reduce gateway timeouts on larger TML imports.
• Fixed import-scope issues in the chunked deploy path.
• Added recovery behavior for ThoughtSpot partial table creation:
  • after 502/503/504 import errors, the deployer checks whether ThoughtSpot actually created the table
  • if the table exists for the same database/schema/connection, the deployer reuses the table GUID instead of failing on table already exists
• Batch 2 join-update failures are now treated as warnings when the run can continue to create a useful model/liveboard.
• Partial liveboard failures are handled more explicitly so a created dataset and model are still shown to the user.

Logging, Run History, And Diagnostics

• Reworked session logging away from shared/module-level state so concurrent or repeated runs are less likely to contaminate each other.
• Session IDs are collision-resistant and include microseconds/random suffixes.
• Added terminal run events:
  • run completed
  • run failed
  • run waiting for user
  • run interrupted
• Run History now distinguishes:
  • Success
  • Failed
  • Partial Success
  • Waiting for User
  • Interrupted
  • Stale / Interrupted
  • No Run Started
• Run History now hides testrunner@thoughtspot.com by default and has a Show test runs toggle.
• Added clearer diagnostics when Snowflake deploy completes but the app does not receive the auto-ThoughtSpot handoff.
• Admin LOG_LEVEL controls logging:
  • off: no Supabase session logging
  • regular: important stage and terminal events
  • verbose: detailed sub-step logs for incident/debug windows

LLM Model Handling

• Fixed Anthropic requests for models that reject deprecated temperature.
• Temporarily forced or defaulted working model paths during provider quota/rate incidents, especially around Sonnet model availability.
• Ensured selected LLM model flows into downstream semantic/model enrichment calls instead of silently using inconsistent defaults.
• Logged model resolution at run start so provider/model issues can be diagnosed from session_logs.

Onboarding And Admin

• Added temporary-password onboarding flow.
• Admin can add users and generate a Slack-ready invite message.
• New users can be forced to change password before accessing the app.
• Removed the unexpected front-end password length requirement.
• Improved onboarding invite copy:
  • welcome message
  • app location
  • docs/quick-start link
  • username and temporary password
  • contact Mike Boone for problems
• Default new-user settings:
  • ThoughtSpot environment: SE Cloud Primary
  • data size: Medium
  • column naming style: Regular Case
  • tag/object prefix/share defaults empty

Test Harness And Deployment Workflow

• The e2e quality harness now resolves schemas from the exact ThoughtSpot model rather than guessing from prefixes or timestamps.
• The harness can reconcile late-completing runs via session_logs.
• Added dataset-first fixed suites and broader randomized/customer-style runs.
• Reconciled GitHub origin/develop and Hugging Face hf-test/main so test and source control are on one code line.
• Documented deploy rule:
  • test: git push hf-test develop:main
  • production only when explicitly requested

Open Follow-Ups

• Date modeling: decide whether most generated models should stop creating noisy physical DATES tables and rely on fact-table date columns plus ThoughtSpot date intelligence instead.
• Liveboards: data scores improved faster than liveboard quality. A separate liveboard redesign effort is in progress.
• ThoughtSpot deployment: continue monitoring 504 and version-conflict behavior under real customer runs.
• Run History: validate the new statuses against real prod/test logs after the next deploy.