Contents

1. Narrow it down in one minute
2. Thalassa only hears me, not the other person
3. Audio arrives but no text appears
4. No answer comes
5. The assistant window is visible in my screen share
6. Answers arrive too late
7. I can't sign in, or I keep getting signed out
8. Minutes or quota used up
9. The answer is in the wrong language
10. A document won't go into the knowledge base
11. Screen analysis does nothing
12. Payment, invoices and cancellation
13. When nothing helps

1. Narrow it down in one minute

Ask these three in order:

1. Is audio arriving? Does the level meter in the assistant window move while someone is speaking? If not, it's an audio problem → section 2.
2. Is audio becoming text? Level moves but no transcript appears? That's speech recognition, and it's almost always the language → section 3.
3. Is text becoming an answer? Transcript is there but no answer? That's the trigger, your quota, or the connection → section 4.

The application distinguishes the first two cases for you. That is built in on purpose: "no signal arrived" and "signal arrived but nothing was recognised" look identical from the outside — nothing happens — but have completely different causes. After roughly twelve seconds without a signal, or twenty without recognition, the app says which of the two it is. That message is the single most useful piece of information in any fault.

What almost never helps: reinstalling. The common causes are a setting in your operating system or in your meeting application, and those survive any reinstall.

2. Thalassa only hears me, not the other person

The most common fault by a wide margin.

On macOS the system does not hand other applications' audio to an app without a one-time setup step. Without it, Thalassa hears only your microphone — which looks exactly like the other person not being recognised. If your own voice appears in the transcript and theirs does not, this is why. The setup step is on the download page.

On Windows Thalassa listens to what the computer plays. If the meeting runs through a different output device than the system default, Thalassa is listening in the wrong place:

1. Make the headset or speaker you actually use the Windows default device.
2. Select the same device in Zoom, Teams or Meet — or "System default".
3. Stop and restart recording in Thalassa.

Second most common on both systems: headphones plugged in or unplugged mid-conversation. The operating system switches device; Thalassa is still on the old one. Stop recording briefly and start again.

What it is not: the meeting platform. Zoom, Teams and Meet need no setup and know nothing about Thalassa. Switching platforms fixes nothing.

3. Audio arrives but no text appears

Usually the transcription language is set to something other than what is being spoken. Choose "Automatic" or the language actually being spoken — you can do this mid-conversation.

Why this happens so often. There are three separate language settings: the interface, the transcript, and the answer. Setting the interface to German does not set the transcription language. If the conversation is in English and the transcript is set to German, recognition finds almost nothing — it is looking for the wrong words.

Changing it mid-conversation is fine. Recognition reconnects, recording keeps running. Stopping and starting is not necessary and would interrupt the session for nothing.

Second cause: too quiet. Audio arrives but stays below the threshold at which it counts as signal — often because the meeting application's volume is set very low. Check both the system volume and the application's.

Third cause: several people talking at once, or heavy background noise. Only the room helps there, not a setting.

4. No answer comes

Four causes, in order of how often they happen:

No question was recognised. Thalassa does not answer every sentence — it answers recognised questions and objections. A question phrased as a statement ("I wonder whether that works with our system") is not always recognised as one. The way around it is built in: you can request an answer yourself at any time, without waiting for recognition.

The quota is empty. Used-up minutes or tier answers are reported. An empty quota does not end a conversation that is already running.

No connection. Answers need the internet; searching your own documents does not. If the app reports it cannot reach the backend, either your connection is down or the service is not reachable.

No access. Without a subscription, credit, ticket or free trial, no conversation starts. You are told before it starts, not halfway through.

What does not happen: a quietly worse answer from a different source. If a model tier fails, you are told — no answer beats an answer of unclear origin.

5. The assistant window is visible in my screen share

Read the indicator. The header of the assistant window states whether concealment is active. If it is not, the window is visible to everyone in the call. That is the one message worth interrupting a conversation over.

The most common reason is not a fault — it is macOS 15.4. From that version, macOS ignores an application's request to be kept out of a share of the whole screen; Apple's developer support has confirmed there is no public interface for it any more. Sharing a single window is not affected. On a current Mac, share the window.

Beyond that, concealment is a request to the operating system — "exclude from capture" on Windows, "do not share" on macOS. If a system prerequisite is missing, the request fails, and the application names what is missing rather than showing a bare error.

What helps, strongest first:

• Use a second display if you have one — move the assistant to a screen you are not sharing. This route depends on no operating-system guarantee at all: what sits on an unshared display is simply not in the recording of the other one. It cannot break, because nothing is promising it. The application suggests this by itself as soon as it sees more than one display.
• Share a single window instead of the whole screen. True on every system.
• Read the message — it names the missing prerequisite.
• Before an important call, check once with a colleague: you share, they tell you whether they can see the window.

A special case that isn't one: anyone photographing or filming your screen with a camera sees the window. Concealment works against screen capture, not against physics.

6. Answers arrive too late

The tiers differ in speed. The deepest tier thinks longer because it does more. If you need quick prompts in conversation rather than a worked-out answer, a faster tier is the most effective lever.

Distance counts. Processing runs in Frankfurt. From Europe you notice nothing; from other continents network travel time is added. That is a property of the EU data path, not a malfunction.

A lot of context costs time. A large case profile, many prepared facts and many hits from your own knowledge base make the request bigger. That is usually worth the time — if you are in a hurry, shorten the preparation.

There is no published latency figure. We do not quote one because none has been reliably measured. An invented number would be wrong exactly when someone checks it.

7. I can't sign in, or I keep getting signed out

The waiting period. Repeatedly wrong passwords cause a pause that grows with each attempt and disappears the moment one succeeds. Your account is not locked; it just takes a moment. If you see a waiting period, reset the password rather than keep guessing.

Google accounts have no password with us. If you originally signed in with Google and are now looking for a password, you are looking for something that never existed. The way in is Google again.

"Invalid or expired access token." Your session expired — access tokens are deliberately short-lived. Signing in again is enough. If it happens constantly, restart the application completely.

Waiting for a confirmation email? None is sent. Registration does not send one and the account works immediately — if you registered, you already have it.

8. Minutes or quota used up

An empty quota does not end a conversation in progress. The only reason a running conversation ends is the plan's fair-use ceiling.

Answers from one tier can run out while minutes remain. Minutes and answer quotas are two separate pots. Answers keep coming, from a different tier — and that is shown.

Subscription minutes do not carry over; purchased minutes expire 365 days after purchase. A balance that suddenly looks low usually has a month boundary behind it.

9. The answer is in the wrong language

Three separate settings — interface, transcript, answer. Changing one does not change the others.

This is the most common confusion in the whole application, and it is deliberate: a French consultant may want a German interface, hold an English conversation, and need German answers. One shared setting would make exactly that case impossible.

Wrong answer language → change the answer language, not the interface. Wrong recognition → change the transcript language.

If translation fails, the language model takes over and the application shows the switch. That is not a fault but a deliberate notice: it is a different processor, and that deserves saying.

10. A document won't go into the knowledge base

The page limit depends on your plan. It counts pages, not files — a single large manual can use it up.

Image-only PDFs. A PDF from a scanner contains pictures of text, not text. There is text recognition for those; it is slower and not equally good on every original. With a poor original, a clean source document beats recognition.

Password-protected files cannot be read. Remove the protection and add it again.

Text recognition stops at 40 pages per document. A case file is not a book, and a 300-page scan would cost more in recognition than it is worth. Split the document, or use a source that already has a text layer.

Very large files take time. Text files are processed on your device, which is why the duration depends on your machine and not on our load.

A scanned PDF is the one exception, and it belongs here rather than in the small print: to recognise text, each page is sent as an image to the backend in Frankfurt. It happens once, at ingest, never during a conversation, and a cache stops the same page being sent twice. If you do not want it, switch text recognition off (KB_OCR=0) — a scan is then reported as an image without a text layer instead of quietly turning into nothing.

11. Screen analysis does nothing

On macOS the screen-recording permission is almost always missing. It is requested at the first attempt and has to be granted in the system dialog: System Settings → Privacy & Security → Screen & System Audio Recording → enable Thalassa. The system does not apply the change to a running application — restart it afterwards. That is a macOS behaviour, not a fault in the app.

Windows requires no separate permission for this.

What screen analysis is and is not: it looks at the screen on request — not continuously, not in the background, not without your involvement.

12. Payment, invoices and cancellation

Payment and subscriptions run through the payment provider, not through us. Invoices, payment methods and cancellation are managed in their customer portal — we have no access to it.

Payment declined. Your bank or card issuer decides that; we only see that it did not work. Try another payment method, or ask your bank.

Paid, but the plan is not active. Activation happens through a callback from the payment provider and usually takes seconds. If it takes longer, sign out and in again; if that does not help, contact support — the purchase is not lost.

Cancelling happens in the payment provider's portal. That it is not with us is not an excuse but a consequence of the contract: the purchase contract is with them. For the same reason, account deletion refuses while a paid subscription is running — cancel first, then delete.

The buy button does nothing. Usually an expired session — sign in again. A browser extension blocking third-party content can also stop checkout.

13. When nothing helps

Write to support@thalassa-consulting.com.

There is no support phone line — the number in the imprint is the legally required contact, not a support line — no 24/7 operation, no ticket system with status tracking, and no guaranteed resolution time.

Four details save the most back-and-forth:

1. Windows or macOS, and which version.
2. Which version of Thalassa.
3. What you did, what you expected, what happened instead.
4. The exact wording of any message that appeared.

The operations log. The application writes technical events to a file in its data folder:

• macOS: ~/Library/Application Support/Thalassa/logs/
• Windows: %LOCALAPPDATA%\Thalassa\logs\

What is not in it: no transcript, no question, no document contents, no credentials. Credentials are stripped not only by field name but also from inside error messages — which is where they end up when nobody is looking. You can therefore send the file without giving up conversation content. Reading it first is still fair.

The log is deliberately unimportant to operation: a full or write-protected disk must never stop a conversation. If the log is missing, only the log is missing.