Conversational Analytics

The chat interface answers plain-language questions with grounded narrative, tables, and charts computed over complete datasets — engineered for numerical accuracy, not a generic chatbot. It adapts its strategy to the size and shape of each result and reconciles what it renders so tables stay consistent across a conversation.

In this guide

Adaptive answering modes

• A mode-decision step measures each result and picks the best strategy: a full-context mode that passes small datasets to the model in full for exact answers, and a data-map mode that summarizes large datasets by schema and sample.
• Tiered deterministic filters run before the model, so filtering, sorting, ranking, and selection are computed in code — not guessed.
• A code-generation path computes aggregates (totals, averages, counts, group-by) deterministically for quantitative questions.

Follow-ups & result reuse

• Filter-only follow-ups reuse the prior result instead of re-running the workflow, for faster responses.
• Underspecified follow-up filters are completed from context; stale carried-over filters are dropped so a new question isn't contaminated by an old one.
• Execution results are rehydrated from cache across turns and reloads, so the conversation stays grounded.

Table & column consistency

• Column selection and display labels are locked and reused, so the same query renders an identical table every time.
• Financial and banking number formats are normalized so the model reasons over clean values.
• Paired-superlative questions ("best and worst") render both tables rather than collapsing to one.

Charts

• On-demand bar, pie, and line charts generated from the answer table, with lazy rendering and category / label projection so charts match the shown data.

File-aware chat

• Attach and query uploaded files — Excel, CSV, DOCX, and text — inside a conversation; native extraction reads file content, and image bytes are reused efficiently across turns.

Accuracy guards & safety

• Empty results return a clear "nothing matched" message instead of a misleading full table; the narrative summary is verified against the final rendered table.
• Every chat request passes the guardrail gate before reaching a model.