AI Assistant & Semantic Search

Most analytics tools make you find the right dashboard, pick the right filters, and hope the answer is in there somewhere. The Astervis assistant flips that around: you type a question the way you'd ask a colleague — "How many calls did we miss yesterday?" — and it answers, with a number, a short explanation, and a chart it picks for you. Underneath, it can also search across what was actually said on your calls, not just the numbers about them.

This page explains what the assistant can do, how it stays safe with your data, and the kinds of questions worth trying first.

Two ways to ask

The assistant answers from two different sources, and it chooses between them (or combines them) based on your question:

Semantic search: find calls by meaning

Keyword search finds the word you typed. Semantic search finds the idea — even when the caller used completely different words. Ask for "customers frustrated about delivery times" and you'll get calls where someone said "my order is late again" or "when is this finally arriving", with no shared keyword between them.

This works because every transcript is turned into a searchable index as part of analysis:

• After a call is transcribed and diarized, each transcript is embedded with Google's gemini-embedding-001 model into 1536-dimension vectors, stored in pgvector right inside your TimescaleDB.
• Transcripts are chunked small-to-big: small child chunks (a few turns each) are what gets matched, while a larger parent segment is returned alongside the match so you get the surrounding context, not a stray sentence.
• Retrieval is hybrid — it blends dense vector similarity (meaning) with keyword (BM25) lexical search (exact terms like a product name or order number) and fuses the two rankings. You get the best of both: a search that understands intent but still nails the literal match when it matters.

The practical payoff: you can ask things like "find calls where someone mentioned a refund" and the assistant surfaces the relevant conversations with a quote and a link straight to each call.

The assistant flow, step by step

When you ask a question, the assistant works through a short, predictable pipeline — and it streams the result back live, so you watch it think rather than staring at a spinner.

Behind the scenes the assistant picks one of three strategies for each question — query the numbers only, search the transcripts only, or query the numbers and then re-rank with transcript search — so a simple count stays fast and a "why" question gets the depth it needs.

Conversations, history and shareable dashboards

The assistant is a place you come back to, not a one-shot box.

• Conversations persist. Follow-up questions keep their context, so you can drill in — "now break that down by operator" — without restating everything.
• Query history is kept. Every question you've asked is saved, so you (or a teammate) can re-run a useful one later instead of rebuilding it.
• Generative dashboards can be saved and shared. When a question produces a chart worth keeping, save it as a dashboard and share it with a link — handy for a recurring number a supervisor or owner wants to glance at without logging into the analytics screens themselves.

Questions worth trying

Start with the kind of thing you'd otherwise build a report for. A few examples:

You don't have to phrase these precisely — the assistant rewrites and clarifies your question before it does anything, and you can always correct it in a follow-up.

Where it fits

The assistant sits on top of everything else the AI pipeline produces, so the richer your transcripts and QA data, the more it can answer.