AI ENGINEER
Citations and fallback
How the AI Engineer shows where each number came from, and what happens when an AI provider is not configured, busy or failing.
The AI Engineer traces every number it reports to a calculation it ran in your project or to data you supplied, and it warns you when a table holds a value with neither source. If the AI provider you chose is not configured, the request fails with a plain message and is never sent elsewhere; if a configured model fails before answering, the AI Engineer moves to another configured model and tells you.
Where its numbers come from
The AI Engineer has three sources, and it keeps them apart:
| Source | How it is shown |
|---|---|
| Calculations it ran in this answer | Model output. Each calculation reports that its numbers were computed by the engine from the model, and that all pressures are absolute, psia. |
| Your messages, pasted text and attachments | Measured data. Quoted as "lab" or "measured", in rows or columns headed Lab or Measured. |
| Assumptions | Stated in words, for example "No C7+ SG was given, so I estimated it from its MW." |
It does not search the web, quote outside documents or recall numbers from memory. It names published methods (for example Whitson gamma splitting or LBC viscosity) because the engine uses them, not as references to look up.
Watching the work
Each calculation appears in the conversation as a step: a spinner while it runs, then a tick (done) or a cross (failed), a short label and, when it finishes, a one-line summary. While it works, a progress line counts the steps and seconds. You can see every calculation behind an answer as it happens.
When a model is created or revised, its card links to it with Open in PVT, so you can check the inputs it read from the report against the source page.
The table check
When an answer finishes, Converge checks every number in every table it contains:
- a number from a calculation in this answer passes,
- a lab number in a row or column labelled Lab or Measured passes,
- a lab number presented as model output is flagged,
- a number with no source is flagged.
Flagged values raise a warning under the answer that starts "Check before you use this answer:", lists up to six values with their row and column, and asks you to treat them as unverified or to ask the AI Engineer to recompute the table from the model.
Pressure basis
Every calculation that takes a pressure needs its basis: psia, psig, bara, barg, kPa or MPa, absolute or gauge. Gauge values are converted to psia before validation or comparison, and the conversion is listed in the result. When a report quoted psig, the AI Engineer gives both, for example "2,449.7 psia (2,435 psig)".
Providers and models
Pick the provider and model in the AI engine panel on the AI Engineer page, or in the side panel. The pickers show model names, not internal ids. Your choice is remembered in your browser.
A provider appears in the AI provider list only as available when the server has it configured. Providers that are not usable are listed but disabled, marked (unavailable), with the reason on hover: "Not configured on this server" or "Busy or out of allowance right now". The AI model list offers only models that are usable now.
When the panel is connected, it reads "Using [provider], [model]."
When the provider is not configured
- No provider configured at all: the provider list reads No AI provider available, an AI Engineer unavailable warning explains why, and the message box says "The AI engine is not available. See the note above."
- You chose a provider that is not configured: the request fails with a clear message. It is never sent to a different provider in its place.
When a model fails
If a configured model fails before it has shown any text or run any calculation, the AI Engineer tries the next model in this order:
- the other available models of the provider you chose,
- then the default and other models of each other provider configured on the server.
A note in the conversation tells you, for example "[model] is not available right now, so [other model] is answering instead." It never falls back to a provider the server has no configuration for.
Fallback happens only before any text or calculation, so a retry cannot repeat work or create duplicate models. Once the answer has started, a failure ends the turn with an error instead.
An authentication failure, a rejected request or a refusal does not trigger a fallback.
How long a failing model is skipped
| Failure | Model hidden from the picker for |
|---|---|
| No usage quota left | 6 hours |
| Rate limited (busy) | 2 minutes, or the provider's own retry time (at least 30 seconds) when it gives one |
| Not responding, server error or network failure | 1 minute |
A model that answers successfully is shown again at once.
Error messages
The AI Engineer shows one plain message per failure. The provider's own error text is logged on the server only.
| Situation | Message |
|---|---|
| Busy or out of allowance | "The AI model [name] is busy or has used its allowance. Try again in a minute, or choose another model." |
| Not responding | "The AI model [name] is not responding right now. Try again in a minute, or choose another model." |
| Server not set up | "The AI service is not set up correctly on the server. Contact support." |
| Request too large | "The AI model [name] could not process this request. Try a shorter message or fewer attachments." |
| Too many steps | "The AI Engineer took too many steps without finishing. Ask a narrower question, or split it into parts." |
| Over 4 minutes | "The AI Engineer did not finish within 240 s and was stopped. Ask a narrower question, or split it into parts." |
After any failure, your message, files and pasted report stay in the box so you can send again.