Files
doctate/server/src/transcribe/worker.rs
T
Brummel d65720c671 Refactor oneliner storage to use enum
The oneliner is now stored as `OnelinerState` enum which can represent
three states: `Ready`, `Empty` or `Error`. This allows the client to
differentiate between a case with no medical content and a case where
the oneliner generation failed.

The `OnelinerState` enum is serialized to JSON and stored in
`oneliner.json` file. The `OnelinerEntry` struct has been updated to
reflect this change.

The client-desktop application has been updated to handle the new
`OnelinerState` enum and display appropriate UI elements for each state.

The server-side code has also been updated to read and write the
`OnelinerState` enum, and to handle the new file format.
The `reset_case_artefacts` function in
`server/src/routes/case_actions.rs` has been updated to remove
`oneliner.txt` and create `oneliner.json` instead.

The tests have been updated to reflect these changes.
2026-04-20 12:37:48 +02:00

448 lines
16 KiB
Rust

use std::path::Path;
use std::sync::Arc;
use std::time::Duration;
use doctate_common::oneliners::{ONELINER_FILENAME, OnelinerState};
use time::OffsetDateTime;
use time::format_description::well_known::Rfc3339;
use tracing::{error, info, warn};
use super::{TranscribeReceiver, ffmpeg, ollama, whisper};
use crate::config::{Config, WhisperUserSettings};
use crate::events::{self, CaseEventKind, EventSender};
use crate::gazetteer::Gazetteer;
use crate::{BusyGuard, WorkerBusy};
const ONELINER_TIMEOUT: Duration = Duration::from_secs(60);
/// Consume transcription jobs sequentially. Whisper holds the GPU, so parallelism
/// here would only cause contention. One job in flight at a time.
pub async fn run(
mut rx: TranscribeReceiver,
config: Arc<Config>,
client: reqwest::Client,
worker_busy: WorkerBusy,
vocab: Arc<Gazetteer>,
events_tx: EventSender,
) {
info!("Transcription worker started");
let timeout = Duration::from_secs(config.whisper_timeout_seconds);
while let Some(job) = rx.recv().await {
let _guard = BusyGuard::new(worker_busy.clone());
let audio_path = job.audio_path;
let transcript_path = audio_path.with_extension("transcript.txt");
if transcript_path.exists() {
continue;
}
// Look up per-user Whisper settings live from the shared config so that
// edits to users.toml (on next restart) reach in-flight jobs naturally.
// Unknown slug → empty settings (service falls back to language=de).
let settings: WhisperUserSettings = config
.users
.iter()
.find(|u| u.slug == job.user_slug)
.map(|u| u.whisper.clone())
.unwrap_or_default();
info!(audio = %audio_path.display(), user = %job.user_slug, "Transcribing");
let remuxed = match ffmpeg::remux_faststart(&audio_path).await {
Ok(f) => f,
Err(e) => {
error!(audio = %audio_path.display(), error = %e, "ffmpeg remux failed");
mark_failed(&audio_path, &events_tx, &job.user_slug).await;
continue;
}
};
// Persist the container duration as a sidecar so the UI can render it
// without the browser having to HEAD every audio file. Non-fatal — the
// recordings scan has a lazy-backfill path, and transcription is the
// critical workload here.
write_duration_sidecar(&audio_path, remuxed.path()).await;
let text = match whisper::transcribe(
&client,
&config.whisper_url,
remuxed.path(),
timeout,
&settings,
)
.await
{
Ok(t) => t,
Err(e) => {
error!(audio = %audio_path.display(), error = %e, "whisper call failed");
mark_failed(&audio_path, &events_tx, &job.user_slug).await;
continue;
}
};
// Normalize terminology against the curated vocab before
// persisting. Downstream consumers (oneliner, analyze) read
// the canonical path.
let text = vocab.replace(&text);
if let Err(e) = tokio::fs::write(&transcript_path, &text).await {
error!(transcript = %transcript_path.display(), error = %e, "writing transcript failed");
continue;
}
info!(
audio = %audio_path.display(),
bytes = text.len(),
"Transcript written"
);
// Notify the live-update bus so subscribed browsers can refresh the
// recordings view. case_id = parent dir of the audio file.
if let Some(case_dir) = audio_path.parent() {
events::emit(
&events_tx,
&job.user_slug,
events::case_id_of(case_dir),
CaseEventKind::TranscriptReady,
);
}
// Regenerate the oneliner, but only after the *last* pending
// recording in the batch — otherwise we'd burn LLM calls on
// incomplete inputs that the next job would redo. Later
// recordings still override earlier ones (symmetric to the
// analyze-LLM "Spätere Aufnahmen haben Vorrang" rule). Silent /
// empty cases are no-ops; LLM errors leave any existing
// oneliner untouched.
if let Some(case_dir) = audio_path.parent()
&& !has_pending_recordings(case_dir).await
{
update_oneliner(
case_dir,
&job.user_slug,
&client,
&config,
&vocab,
&events_tx,
)
.await;
}
}
warn!("Transcription worker stopped (channel closed)");
}
/// Atomically rename a failed recording from `<ts>.m4a` to `<ts>.m4a.failed`
/// so the recovery scan no longer picks it up. The UI still surfaces these
/// files (with a "failed" flag) so the user can listen to the audio and
/// manually decide whether to retry (by renaming back) or to delete.
async fn mark_failed(audio_path: &Path, events_tx: &EventSender, user_slug: &str) {
let failed_path = audio_path.with_extension("m4a.failed");
if let Err(e) = tokio::fs::rename(audio_path, &failed_path).await {
// Nothing to roll back — if rename fails, the worst case is that we
// retry the same file on next restart. Log and move on.
error!(
audio = %audio_path.display(),
error = %e,
"Failed to mark recording as failed; may be retried on restart",
);
} else {
warn!(
audio = %audio_path.display(),
failed = %failed_path.display(),
"Recording marked as failed",
);
if let Some(case_dir) = audio_path.parent() {
events::emit(
events_tx,
user_slug,
events::case_id_of(case_dir),
CaseEventKind::TranscriptFailed,
);
}
}
}
/// Probe the (remuxed) audio and persist the duration next to the original file
/// as `<ts>.duration.txt`. Probes on the remuxed copy because its `moov` atom
/// sits at the start — ffprobe returns without seeking to EOF.
async fn write_duration_sidecar(audio_path: &Path, probe_path: &Path) {
let secs = match ffmpeg::probe_duration_seconds(probe_path).await {
Ok(s) => s,
Err(e) => {
warn!(audio = %audio_path.display(), error = %e, "ffprobe duration failed — UI will lazy-backfill");
return;
}
};
let sidecar = audio_path.with_extension("duration.txt");
if let Err(e) = tokio::fs::write(&sidecar, secs.to_string()).await {
warn!(sidecar = %sidecar.display(), error = %e, "writing duration sidecar failed");
}
}
/// Regenerate `case_dir/oneliner.json` from **all** non-empty transcripts
/// in the case, joined chronologically. Called after every successful
/// transcript write so later recordings can correct earlier ones (mirror
/// of the analyze-LLM rule "later recordings override"). Every outcome —
/// generated text, deliberate empty ("no medical content"), or call error
/// — is persisted as an [`OnelinerState`] variant, overwriting any
/// previous state. Silent case (no non-empty transcript) is a no-op; the
/// next transcript write retries.
pub(crate) async fn update_oneliner(
case_dir: &Path,
user_slug: &str,
client: &reqwest::Client,
config: &Config,
vocab: &Gazetteer,
events_tx: &EventSender,
) {
let transcript = match all_transcripts_joined(case_dir).await {
Some(t) => t,
None => return,
};
let generated_at = OffsetDateTime::now_utc()
.format(&Rfc3339)
.unwrap_or_default();
let state = match ollama::generate_oneliner(
client,
&config.ollama_url,
&config.ollama_model,
config.ollama_keep_alive,
&transcript,
ONELINER_TIMEOUT,
)
.await
{
Ok(line) => {
let text = vocab.replace(&line);
info!(
user = %user_slug,
case = %case_dir.display(),
chars = text.chars().count(),
"Oneliner ready"
);
OnelinerState::Ready { text, generated_at }
}
Err(ollama::OllamaError::EmptyResponse) => {
warn!(
case = %case_dir.display(),
"oneliner empty — model returned nothing (no medical keyword in transcript)"
);
OnelinerState::Empty { generated_at }
}
Err(e) => {
error!(case = %case_dir.display(), error = %e, "oneliner generation failed");
OnelinerState::Error { generated_at }
}
};
if let Err(e) = write_oneliner_state(case_dir, &state).await {
error!(case = %case_dir.display(), error = %e, "writing oneliner.json failed");
return;
}
events::emit(
events_tx,
user_slug,
events::case_id_of(case_dir),
CaseEventKind::OnelinerUpdated,
);
}
/// Write `state` atomically to `case_dir/oneliner.json` via
/// `write(tmp) + rename(tmp, final)`. Rename is atomic within a single
/// filesystem, so concurrent readers never see a half-serialized JSON
/// document.
async fn write_oneliner_state(case_dir: &Path, state: &OnelinerState) -> std::io::Result<()> {
let final_path = case_dir.join(ONELINER_FILENAME);
let tmp_path = case_dir.join(format!("{ONELINER_FILENAME}.tmp"));
let payload = serde_json::to_vec(state).expect("OnelinerState is infallibly serializable");
tokio::fs::write(&tmp_path, &payload).await?;
tokio::fs::rename(&tmp_path, &final_path).await
}
/// Return true if `case_dir` contains at least one `.m4a` recording
/// that has no corresponding `.transcript.txt`. Used to gate the
/// oneliner-regenerate call: we only run it after the *last* pending
/// transcription in a batch, not after every single one. `.m4a.failed`
/// entries are skipped — they don't block the oneliner. Fail-open on
/// read errors: if we can't scan, return false so the oneliner runs
/// anyway (one redundant call beats never).
pub(crate) async fn has_pending_recordings(case_dir: &Path) -> bool {
let Ok(mut entries) = tokio::fs::read_dir(case_dir).await else {
return false;
};
while let Ok(Some(entry)) = entries.next_entry().await {
let path = entry.path();
if path.extension().and_then(|s| s.to_str()) != Some("m4a") {
continue;
}
if !path.with_extension("transcript.txt").exists() {
return true;
}
}
false
}
/// Read every non-empty `*.transcript.txt` in `case_dir`, ordered
/// chronologically (ISO-8601 filename sort equals chronological order), and
/// join them with `\n\n---\n\n`. Empty / whitespace-only transcripts (silent
/// recordings) are skipped. Returns None if no non-empty transcript exists.
pub(crate) async fn all_transcripts_joined(case_dir: &Path) -> Option<String> {
let mut entries = tokio::fs::read_dir(case_dir).await.ok()?;
let mut files: Vec<std::path::PathBuf> = Vec::new();
while let Ok(Some(entry)) = entries.next_entry().await {
let name = entry.file_name();
let Some(s) = name.to_str() else { continue };
if s.ends_with(".transcript.txt") {
files.push(entry.path());
}
}
files.sort();
let mut parts: Vec<String> = Vec::with_capacity(files.len());
for path in files {
if let Ok(text) = tokio::fs::read_to_string(&path).await {
let trimmed = text.trim();
if !trimmed.is_empty() {
parts.push(trimmed.to_string());
}
}
}
if parts.is_empty() {
return None;
}
Some(parts.join("\n\n---\n\n"))
}
#[cfg(test)]
mod tests {
use super::*;
use tempfile::tempdir;
/// Regression guard for the Womax bug: the joined transcript must
/// include ALL recordings in chronological order, so the LLM sees
/// both the initial term and any later correction.
#[tokio::test]
async fn all_transcripts_joined_includes_all_in_order() {
let dir = tempdir().unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-05-00Z.transcript.txt"),
"Korrektur. Das Mittel heißt Vomex.",
)
.await
.unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-00-00Z.transcript.txt"),
"Jetzt weiß ich es wieder. Es heißt Womax.",
)
.await
.unwrap();
let got = all_transcripts_joined(dir.path()).await.expect("Some");
assert_eq!(
got,
"Jetzt weiß ich es wieder. Es heißt Womax.\n\n---\n\nKorrektur. Das Mittel heißt Vomex."
);
}
/// Silent recordings are skipped, not treated as end-of-stream — a
/// later non-empty recording still contributes.
#[tokio::test]
async fn all_transcripts_joined_skips_empty_recordings() {
let dir = tempdir().unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-00-00Z.transcript.txt"), "")
.await
.unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-05-00Z.transcript.txt"),
"non-empty content",
)
.await
.unwrap();
let got = all_transcripts_joined(dir.path()).await.expect("Some");
assert_eq!(got, "non-empty content");
}
#[tokio::test]
async fn all_transcripts_joined_is_none_when_all_empty() {
let dir = tempdir().unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-00-00Z.transcript.txt"),
" \n ",
)
.await
.unwrap();
assert!(all_transcripts_joined(dir.path()).await.is_none());
}
#[tokio::test]
async fn all_transcripts_joined_is_none_when_no_transcripts() {
let dir = tempdir().unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-00-00Z.m4a"), b"audio")
.await
.unwrap();
assert!(all_transcripts_joined(dir.path()).await.is_none());
}
#[tokio::test]
async fn has_pending_recordings_true_when_m4a_without_transcript() {
let dir = tempdir().unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-00-00Z.m4a"), b"a")
.await
.unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-00-00Z.transcript.txt"),
"text",
)
.await
.unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-05-00Z.m4a"), b"b")
.await
.unwrap();
assert!(has_pending_recordings(dir.path()).await);
}
#[tokio::test]
async fn has_pending_recordings_false_when_all_transcribed() {
let dir = tempdir().unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-00-00Z.m4a"), b"a")
.await
.unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-00-00Z.transcript.txt"),
"text1",
)
.await
.unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-05-00Z.m4a"), b"b")
.await
.unwrap();
tokio::fs::write(
dir.path().join("2026-04-16T10-05-00Z.transcript.txt"),
"text2",
)
.await
.unwrap();
assert!(!has_pending_recordings(dir.path()).await);
}
#[tokio::test]
async fn has_pending_recordings_ignores_failed() {
let dir = tempdir().unwrap();
tokio::fs::write(dir.path().join("2026-04-16T10-00-00Z.m4a.failed"), b"x")
.await
.unwrap();
assert!(!has_pending_recordings(dir.path()).await);
}
#[tokio::test]
async fn has_pending_recordings_false_for_empty_dir() {
let dir = tempdir().unwrap();
assert!(!has_pending_recordings(dir.path()).await);
}
}