Session: prior question and choice
Semantics: metric and term definition
Identity: project and data scope
Time: cutoff and comparison window
Do not treat the transcript as state
When a user says “what about South China?”, the system must resolve the prior metric, comparison period, filters, and the user’s access. A long transcript may contain words but not a reliable structure.
Represent context as metric, dimensions, filters, time, identity, and output format. Record each change rather than silently rewriting the whole query.
Separate inherited and revalidated values
Metric and comparison period can usually carry forward. Project changes, sensitive fields, or a new data cutoff should trigger a check. “Use this year instead” should change only the time slot.
Test omitted subjects, pronouns, corrections, topic switches, and recovery of a saved conversation.
Bind context to governance
A context object must include user, organization, project, and policy version. A shared cache keyed only by question text can leak one role’s result to another.
AskTable.ai confirms organization, project, and data-scope controls. Conversation persistence, cache policy, and IAM integration still need confirmation for each deployment.
Log context for review
Store state, user edits, final conditions, and result summaries so a reviewer can see which turn changed the definition. Expire old sessions to reduce contamination from obsolete metric rules.
Public references
Ready to help your team start?
Talk through a real scenario and see how AskTable.ai can fit your business.
Book a demo