# Search and ask across a client's meetings

> Inside a client, ask the assistant a question in plain language and get an answer with citations to who said it and when, linked to the moment in the recording. Across clients, one search box finds meetings, action items, summaries and transcript lines, each labelled with its client. For anyone who records client calls on Google Meet or Microsoft Teams.

Canonical: https://trackheros.com/features/ask-your-meetings/  
Updated: 2026-09-02

Every meeting TrackHeros records on Google Meet or Microsoft Teams becomes something you can search and question. Inside a client, you ask in plain language — "what did we agree about the API deadline?" — and get an answer with citations to who said it and when, linked to that moment in the recording. Across clients, one search box finds meetings, action items, summaries and transcript lines, each labelled with the client it belongs to. This page is for anyone who has scrolled a transcript at midnight looking for a promise they made three weeks ago.

## What does it do?

- Answers questions from one client's meetings, with a citation on every claim.
- Searches everything you own — clients, meetings, action items, summaries, transcripts — keyword first, with semantic matches shown separately.
- Inside a client, acts on your behalf: corrects a summary, edits or resolves an action item, creates a correction rule, re-summarises a meeting. Every change is shown as a plan you approve before it runs.
- Offers a global, read-only chat mode across your own clients, with every fact labelled by client.
- Takes dictation into the composer and exports any answer as Markdown, PDF or email HTML.
- No credits, no limits. Ask as much as you want.

## How does it work?

| Where you are | What it reads               | What it can do                                                     |
| ------------- | --------------------------- | ------------------------------------------------------------------ |
| Client chat   | That client's meetings only | Answer with citations; propose changes for your approval           |
| Global search | Every active client you own | Return rows, each labelled with its client; no answer is generated |
| Global chat   | Every active client you own | Answer with each fact labelled by client; change nothing           |

An answer is not a paragraph you have to take on trust. Each claim carries a citation: the meeting, the speaker and the time. Open it and the recording plays from that moment, inside the app, through a short-lived link tied to the meeting. There are no public recording URLs. Where the transcript could not attribute a line to a person, the citation says Unknown rather than picking the likeliest name.

The assistant retrieves before it answers. Inside a client, the retrieval is scoped to that client before it runs, which is why it cannot be talked into reading another one: the query never fetched those rows. How that scoping is enforced is on the [client isolation page](/features/client-isolation/).

## What can the assistant change, and how is that kept safe?

Inside a client, the assistant is allowed to do four things beyond answering:

1. Correct a summary: a wrong name, a missed decision, a mangled term.
2. Edit or resolve an action item: change the owner, fix the due date, mark it done.
3. Create a correction rule for the client.
4. Re-summarise a meeting, for example after a rule was added or to switch it to the interview profile.

Before any of them, it shows a plan: what it intends to change, on which meeting, from what to what. Then it waits. Nothing runs until you approve; decline and nothing happens. There is no setting that lets it act first and ask later. The rule is the same for every action, every time: you approve every change, one by one.

## How does global search work?

One box over everything you own. It has two arms, and they are kept apart on purpose.

Keyword is the primary arm: a plain text match across client names, meeting titles and invitation text, action items (task, owner, due phrase), summary lines and transcript lines. Results come back grouped by kind, narrowest first — a client is one row that reframes the search, a transcript line is the deepest thing held — and newest first within a group.

Semantic is the second arm: rows whose meaning matches, shown underneath as a separate, labelled group of suggestions. Anything keyword already found is removed from it, so the second group reads as "and also these" rather than a reshuffle, and it has a distance ceiling so a typo does not come back as the least-unrelated rows in your account. Semantic is on by default and can be turned off.

A few deliberate limits. Search is a search box, not the assistant: no model reads the results and no answer is generated from them; you get rows and click one. Results from a single long meeting are capped so one chatty call cannot fill the page. Archived clients are left out unless you name one. And every row carries its client, which is the structural half of "never blend".

## What is the global chat mode?

A read-only assistant over your own history across clients. It answers a question like "which clients have an open item due this week?" with a list of facts, each labelled with its client. It never writes one narrative that draws on two clients' content, and if a question would need that, it does not do it. It also changes nothing: no summary edits, no action items, no rules. To change something, enter the client.

## What are correction rules?

Two kinds, both per client:

| Kind        | Example                                                      | Effect                                                       |
| ----------- | ------------------------------------------------------------ | ------------------------------------------------------------ |
| Vocabulary  | "UBVA" is a mishearing of "UBA"                              | The transcript and the summary always write the correct form |
| Instruction | "Ticket ids always take the form PROJ- followed by a number" | The convention is applied when transcribing and summarising  |

Rules apply to every future meeting for that client. A past meeting is fixed by re-summarising it. You can create a rule by telling the assistant ("from now on write UBA, not UBVA"), which shows the plan and waits for approval, or from "Propose a correction" on a meeting page. How rules feed the minutes is on the [meeting minutes page](/features/meeting-minutes/).

## Can I dictate, and can I take an answer with me?

Dictation: press the microphone in the composer and speak. The audio goes through the same transcription pipeline as your meetings and comes back as text in the composer, for you to edit before sending. Nothing is stored; the audio is not kept and the dictation is not saved anywhere.

Export: any answer can be taken out as Markdown, as a PDF, or as email-ready HTML to paste into a message to a client.

There are no credits to spend and no cap on questions. A [freelance developer](/for/freelance-developers/) with six clients can ask each workspace about every meeting it holds, as often as needed.

## Frequently asked

**Does the assistant read all my clients' meetings when I ask a question?**

No. Inside a client, it retrieves from that client's meetings only. The global chat mode reads across your own clients, but it labels every fact with its client, never blends them into one narrative, and cannot change anything.

**Can the assistant change my notes?**

Inside a client, yes, with your approval. It can correct a summary, edit or resolve an action item, create a correction rule, or re-summarise a meeting, and each of those is shown to you as a plan before it runs. Nothing changes until you approve it.

**What is the difference between search and the assistant?**

Search returns rows: meetings, action items, summary and transcript lines, each labelled with its client, and no answer is generated from them. The assistant reads the retrieved passages and writes an answer with citations. Search is for finding the thing; the assistant is for asking about it.

**Are there limits on how many questions I can ask?**

No. There are no credits to spend and no cap on questions. Ask as much as you want, in any client, and export any answer as Markdown, PDF or email HTML.

---

TrackHeros is in private beta. Early access: https://trackheros.com/#early-access