Quick analysis: one question, one answer

Step 1, connect a model

1. Ask, and skip this step entirely

   Press Analyze on any recording, choose Quick analysis, and type a question. With webQsee AI there is nothing to set up first: no settings form opens, because there is no key to paste and no address to find. Your plan includes a number of analyses each month.

   Sanity check: the panel opens straight on the question box, and the line above it says the excerpt goes to webQsee.

2. Switch the engine without leaving the dialog

   Selecting the Quick analysis card opens a Which AI answers? row inside it, grouped by who pays: webQsee AI with its three strengths, and under BYO AI whatever you have configured. Each entry says what it is and what is left of it, and picking one repaints the card in that engine's colour, so the choice is made and visible before you press Start. You only need the settings if the engine you want is not configured yet, and the card has its own button for that.

   Sanity check: the card's chip at the top changes with your pick, from amber webQsee AI to cyan BYO AI or the other way round. Colour is never the only signal: the chip names the engine too.

3. Or open the settings and bring your own model or agent

   Quick analysis settings on the card, or the gear icon in the analysis panel, opens the settings over the dialog rather than in a new tab, and what you change there is in the card the moment you close it. Three choices, and which one you want depends on whether the recording may leave your machine at all.

   • webQsee AI, in three strengths. Nothing to configure. The excerpt goes to webQsee and on to a model; webQsee counts the analysis and keeps no part of the request. Standard is included in your plan up to a monthly number and is enough for most recordings. A free account gets five Standard analyses on the house, once, so that trying it costs nothing and needs no key; the panel offers them the first time you have none left; Smart and Genius are bought as boosters: a stronger model, and the largest we host, for escalating a question Standard did not settle. You pick a strength on the cards at the top of the settings, per analysis, and each card shows what you have left.
   • Anthropic. Paste an API key. The default model is a current Claude model.
   • OpenAI-compatible endpoint. Any server speaking that API: Ollama, LM Studio, vLLM, a gateway of your own, or a hosted provider. You give the base URL and the model name.
   Sanity check: for an OpenAI-compatible endpoint the URL ends in /v1, for example http://localhost:11434/v1.

4. Decide where the recording is allowed to go

   An address on your own machine keeps the recording on your own machine. The settings form says so under the endpoint field, and the sentence above the question box repeats it every time you ask, because "runs locally" and "is posted to a provider" are two clicks apart and look identical otherwise. With webQsee AI that sentence names webQsee, because with that option the destination is us.

   Sanity check: the line above the question box names the destination you expect, with a shield icon if masking is on and a warning triangle if it is not.

5. Save

   The API key is stored in the browser's own extension storage and never reaches the webQsee backend, the helper, or anything else. It is sent to the endpoint you named and nowhere else.

   Sanity check: the settings form closes and the question box is enabled.

Step 2, ask well

The quick analysis sees a selection, not everything. What you ask decides how much of that selection is useful, so name the moment and the symptom rather than asking for a summary.

Step 3, read the coverage box

Above the answer there is a box that says what the model was actually given: how many events were in scope, how many response bodies and video frames were sent, and what was dropped to stay inside the token budget. Read it before you read the answer.

It also states whether your own error-detection rules took part. There are three distinct states and they mean different things:

• Your rules were applied. Some verdicts in the answer are yours, not webQsee's.
• You have no rules configured. Nothing was judged by them, and that is not a gap.
• Your rules could not be reached from this surface. That one is worth acting on.

Step 4, follow up and attach

The conversation stays open. A follow-up replays the same excerpt, so the model still has the material and you are not paying to re-send it in a new conversation.

If the answer needs something the excerpt did not include, attach it: pick an extra response body or an extra video frame from the picker and it travels with your next question. A meter shows what the attachment costs against your token budget, counting the conversation so far.

Step 5, keep it

Press Save with the recording and it is stored inside the recording, together with the conditions it was made under: which model answered, whether masking was on, whether your own rules took part. Those conditions are what make a finding reproducible later, which is why they travel with it. Name it in the field beside the button if you want the stored list to lead with something other than your first question.

The row below offers Markdown, Jira markup and Print / PDF, for the copy that goes to somebody who does not have webQsee. Both rows wait until the answer is in.

Troubleshooting

Next