Chat & answers

Ask in plain language; get answers that are grounded, computed, and reasoned, with the work shown.

The answer pipeline

Every message runs through three layers:

1. Layer 1 · RouteA single classification decides everything downstream: the domain (financial, technical, legal, HR, sales, data, or general), the best source (documents, SQL tables, or both), and the difficulty (direct vs. multi-step). Routing is catalog-aware: it knows which tables your workspace actually has.
2. Layer 2 · ExpertThe matching specialist answers in the right voice with the right contract. The financial analyst reports currency, units, and period and shows derivations. The technical expert answers API questions with complete, runnable curl commands. Legal quotes the governing clause verbatim; HR is warm and precise; the data analyst states the result first, then exactly how it was derived.
3. Layer 3 · ReasonHard questions are decomposed into sub-questions, each researched independently (retrieval or SQL), then synthesized into one answer. The plan and each step stream to the UI as they execute.

Citations you can click

Answers grounded in documents carry numbered citations. Clicking one opens the source: the exact passage highlighted in the document, or the exact rows of a spreadsheet. If retrieval finds only weak evidence, the pipeline re-routes to a better source (for example, from documents to SQL for a quantitative question) rather than composing over thin support.

Numbers computed with SQL

Quantitative questions are never answered by language-model arithmetic. The pipeline writes a read-only SQL query over your ingested tables, executes it, and composes the answer from the actual result set. Every computed figure carries a SQL chip that opens the query.

You:  Compare average machine uptime between our sites.

SQL:  SELECT site, AVG(uptime_pct) AS avg_uptime
      FROM machine_telemetry
      GROUP BY site ORDER BY avg_uptime DESC

A:    Berlin averages 96.2% uptime vs Pune at 95.9%.
      Berlin is ahead by 0.3 points. [SQL]

• Guarded: queries are read-only, injection-checked, and row-capped.
• Comparative questions return all groups. A "which site is better" question shows every site, not just the winner.
• Top-K works as expected. "Top 3 customers" returns exactly three, ranked.

Visual mode

Chat and Visual are two views of the same conversation. Visual is a full-screen, animated stage that shows what's actually happening: chunks being retrieved, the route being chosen, sub-questions fanning out, the answer assembling. Use it to build intuition, to debug a surprising answer, or to demo the product.

Voice

Ask by voice in the browser and get spoken answers through the realtime pipeline. Voice answers go through the same grounding; they're the same cited, computed answers, spoken.

Models

Vectropic presents one tuned model surface, Vectropic 1.0, so your team isn't choosing between vendor model names. Additional models will appear here as they're added.