Live report fill
Live report fill is the core of GigHz Precision AI: the report builds itself while you dictate, so by the time you stop talking most of the work is done. This page describes the behaviour you will see and the reasoning behind it.
The skeleton appears when you name the study
As soon as the transcript names a study, the matching template appears as a skeleton: every section and line present, each organ line showing its standard normal. The skeleton is the contract the placement engine works against. If the study is named more precisely later ("with and without contrast", "CT urogram"), the skeleton is swapped for the right variant and everything already placed is re-placed into it.
One sentence, one placement
Each transcribed sentence is sent to a placement model together with the current state of the report: every line, its label, whether it is still the default normal or already holds a dictated finding, and the last two sentences placed. The model returns a small set of operations: replace a line's default with the finding, append an additional finding to a line that already holds one, fill a bracketed blank, or file the sentence under Additional findings. Those operations are applied deterministically on your screen.
Rules the engine follows, which you can rely on:
- A dictated finding always replaces the default normal for its line. "Liver: unremarkable" becomes your finding, never both.
- A second finding for the same organ accumulates. Dictating a second lung nodule adds a second line under Lungs; it never overwrites the first.
- Broad findings go on the broad line. "Multilevel degenerative changes" lands on the spine line once, not fanned across every level.
- Named vessels go on their vessel line, so a celiac stenosis dictated after an SMA finding joins it on the mesenteric line instead of being parked.
- Nothing is dropped. A sentence the engine cannot place is kept verbatim under Additional findings with a review flag.
Pauses, fragments and the transcriber's periods
The live transcriber ends a segment whenever you pause and starts the next one with a capital letter. A breath in the middle of a finding therefore arrives as two "sentences": "7 mm right." and "Splenic lobe cyst." GigHz Precision AI does not trust those breaks. A segment whose tail cannot end a radiology sentence (a conjunction, a preposition, a bare laterality, a measurement with no noun) is held for a moment and joined to what follows; a segment that starts in lowercase or with a conjunction is treated as a continuation. The joined sentence is placed once, as one finding. A period you actually dictate is always honoured.
It keeps working when the tab is in the background
Radiologists dictate with PACS in front and the browser behind it. Browsers slow the timers of a hidden tab to once a minute, which would stall placement until the tab is brought forward. GigHz Precision AI runs its placement timers in a background worker that browsers do not throttle, so placement continues at full speed while you look at the images.
Editing the transcript: rebuild, not patch
If you edit the transcript (not just add to it), the report on screen still reflects the original wording and the old placements cannot be unpicked line by line. The application detects the divergence and offers Rebuild report from edited transcript: the skeleton is reset, every line of the edited transcript is placed again live, and the progress counter runs as it did the first time. Appending new sentences never needs a rebuild; they are placed incrementally.
Finalizing
Finalize waits for any sentence still in flight, composes the impression from the placed findings (see Impression policy), and locks the report. A completeness check runs at this moment: every abnormal finding you dictated must appear somewhere in the report, and anything missing is added back. Unfinalize reopens the live report so you can keep dictating into it.
What happens if the placement service is unavailable
If the placement model cannot be reached, dictation is still captured and every sentence is parked under Additional findings with a visible warning. In that state the application deliberately does not compose an impression from unplaced text: the impression shows one honest line pointing at the parked findings, so a service outage can never produce a confident-looking but wrong impression.
This page is generated from documentation that ships with the application (docs v1.2.2). Version numbers, prices and trial lengths on this page are read from the running system, not typed by hand.