M3c: meeting summaries and agenda coverage via Ollama

- meetrec-server: Ollama integration (chat API, gemma4:12b, num_ctx
  32768); German structured summary (topic/points/decisions/to-dos)
  written to summary.md; agenda coverage returns strict JSON (covered,
  time, evidence) parsed defensively; both run automatically in a
  background thread after upload plus manual trigger endpoints
  (POST /summary, POST /agenda) with status tracking in the index
- phone: agenda input on the Record tab (one item per line, persisted)
  is uploaded with the recording; Library detail shows the summary and
  a per-item agenda checklist with timestamps and evidence quotes,
  with polling while the server generates and manual re-trigger buttons
- validated end-to-end against the live Ollama server: crafted German
  test meeting produced a correct structured summary and perfect agenda
  discrimination (covered items with correct timestamps + quotes,
  undiscussed item correctly false)
This commit is contained in:
2026-09-08 12:32:20 +02:00
parent 4a2aeb60d8
commit 2cf785746e
6 changed files with 385 additions and 9 deletions
+15
View File
@@ -37,6 +37,21 @@ The **Library** tab browses everything:
the server or from the local file), share, and delete with
confirmation. Refresh reloads both lists.
## Summaries and agenda (Ollama)
Enter **agenda items on the Record tab (one per line, persisted)** before
recording. After the recording uploads, meetrec-server automatically:
1. writes a German **summary** (gemma4:12b via your Ollama server):
topic, key points, decisions, to-dos — shown on the recording's
detail page
2. checks **which agenda items were actually discussed** — the detail
page shows a ✓/✗ checklist with the timestamp and a quote as evidence
Both can be re-triggered from the detail page ("Neu erstellen" /
"Agenda neu prüfen"). Note: very long meetings may exceed the LLM's
context window (the transcript is then truncated).
## Requirements
- Android Studio (or: SDK Platform 36, Build Tools 36, NDK 27.1, CMake 3.22.1)
@@ -33,6 +33,7 @@ import androidx.compose.ui.platform.LocalContext
import androidx.compose.ui.unit.dp
import java.io.File
import kotlinx.coroutines.Dispatchers
import kotlinx.coroutines.delay
import kotlinx.coroutines.launch
import kotlinx.coroutines.withContext
import org.json.JSONObject
@@ -177,6 +178,8 @@ private fun LibraryDetail(
var detail by remember { mutableStateOf<Detail?>(null) }
var detailError by remember { mutableStateOf<String?>(null) }
var confirmDelete by remember { mutableStateOf(false) }
var meta by remember { mutableStateOf((item as? LibraryItem.Server)?.meta) }
var summaryText by remember { mutableStateOf<String?>(null) }
// simple streaming player for the WAV
var playing by remember { mutableStateOf(false) }
@@ -202,6 +205,38 @@ private fun LibraryDetail(
}
}
// poll while the server is generating (summary + agenda run via Ollama)
LaunchedEffect(meta?.summary, meta?.agendaStatus) {
val m = meta ?: return@LaunchedEffect
val url = storageUrl.trim().trimEnd('/')
if (m.summary == "pending" || m.agendaStatus == "pending") {
delay(5_000)
try {
meta = withContext(Dispatchers.IO) {
StorageClient.fetchMeta(url, m.id)
}
} catch (_: Exception) {
}
}
}
// fetch the summary once it is marked done
LaunchedEffect(meta?.summary) {
val m = meta ?: return@LaunchedEffect
val url = storageUrl.trim().trimEnd('/')
summaryText = if (m.summary == "done") {
try {
withContext(Dispatchers.IO) {
StorageClient.fetchFile(url, m.id, "summary.md").decodeToString()
}
} catch (e: Exception) {
"(summary unavailable: ${e.message})"
}
} else {
null
}
}
LaunchedEffect(item) {
detail = null
detailError = null
@@ -339,6 +374,86 @@ private fun LibraryDetail(
)
}
// summary + agenda (server recordings only)
meta?.let { m ->
val url = storageUrl.trim().trimEnd('/')
Text("Zusammenfassung", style = MaterialTheme.typography.titleSmall)
when {
m.summary == "pending" -> Text("wird erstellt …",
style = MaterialTheme.typography.bodySmall)
summaryText != null -> Text(summaryText!!,
style = MaterialTheme.typography.bodyMedium)
m.summary != null && m.summary.startsWith("error") -> {
Text(m.summary, color = MaterialTheme.colorScheme.error,
style = MaterialTheme.typography.bodySmall)
}
}
OutlinedButton(onClick = {
scope.launch {
try {
withContext(Dispatchers.IO) {
StorageClient.triggerSummary(url, m.id)
}
meta = withContext(Dispatchers.IO) {
StorageClient.fetchMeta(url, m.id)
}
} catch (e: Exception) {
detailError = "summary trigger failed: ${e.message}"
}
}
}) {
Text(if (m.summary == null) "Zusammenfassung erstellen" else "Neu erstellen")
}
if (m.agenda.isNotEmpty()) {
Text("Agenda", style = MaterialTheme.typography.titleSmall)
if (m.agendaStatus == "pending") {
Text("wird geprüft …", style = MaterialTheme.typography.bodySmall)
}
m.agendaResults?.forEach { r ->
if (r.error != null) {
Text("Agenda-Prüfung fehlgeschlagen: ${r.error}",
color = MaterialTheme.colorScheme.error,
style = MaterialTheme.typography.bodySmall)
} else {
Row(horizontalArrangement = Arrangement.spacedBy(6.dp)) {
Text(
if (r.covered) "" else "",
color = if (r.covered) {
MaterialTheme.colorScheme.primary
} else {
MaterialTheme.colorScheme.error
},
)
Text(
r.item + if (r.time.isNotBlank()) " (${r.time})" else "",
style = MaterialTheme.typography.bodyMedium,
)
}
if (r.evidence.isNotBlank()) {
Text("\u201C${r.evidence}\u201D",
style = MaterialTheme.typography.bodySmall,
color = MaterialTheme.colorScheme.onSurfaceVariant)
}
}
}
OutlinedButton(onClick = {
scope.launch {
try {
withContext(Dispatchers.IO) {
StorageClient.triggerAgenda(url, m.id)
}
meta = withContext(Dispatchers.IO) {
StorageClient.fetchMeta(url, m.id)
}
} catch (e: Exception) {
detailError = "agenda trigger failed: ${e.message}"
}
}
}) { Text("Agenda neu prüfen") }
}
}
detailError?.let {
Text(it, color = MaterialTheme.colorScheme.error,
style = MaterialTheme.typography.bodySmall)
@@ -356,7 +471,7 @@ private fun LibraryDetail(
} else {
Text("Transcript (${d.segments.size} segments)",
style = MaterialTheme.typography.titleSmall)
LazyColumn(Modifier.fillMaxSize()) {
LazyColumn(Modifier.fillMaxWidth().weight(1f)) {
items(d.segments) { (startMs, text) ->
Column(Modifier.fillMaxWidth().padding(vertical = 4.dp)) {
Text(fmtMs(startMs), style = MaterialTheme.typography.labelSmall)
@@ -76,6 +76,7 @@ fun RecordScreen() {
?: "http://100.103.83.12:8090",
)
}
var agendaText by remember { mutableStateOf(prefs.getString("agenda", "") ?: "") }
LaunchedEffect(transcribeOn) {
prefs.edit().putString("transcribe_on", transcribeOn).apply()
@@ -86,6 +87,9 @@ fun RecordScreen() {
LaunchedEffect(storageUrl) {
prefs.edit().putString("storage_url", storageUrl).apply()
}
LaunchedEffect(agendaText) {
prefs.edit().putString("agenda", agendaText).apply()
}
var model by remember { mutableStateOf("tiny") }
var modelMenu by remember { mutableStateOf(false) }
@@ -207,6 +211,7 @@ fun RecordScreen() {
null
},
storageUrl = storageUrl.trim().ifBlank { null },
agenda = agendaText.lines().map { it.trim() }.filter { it.isNotBlank() },
)
if (transcribeOn == "server") {
status = "Recording — live and final pass on ${serverUrl.trim()}"
@@ -301,6 +306,17 @@ fun RecordScreen() {
modifier = Modifier.fillMaxWidth(),
)
// agenda: one item per line, checked against the transcript after upload
OutlinedTextField(
value = agendaText,
onValueChange = { agendaText = it },
label = { Text("Agenda (one item per line)") },
placeholder = { Text("Budget\nZeitplan\n") },
minLines = 1,
maxLines = 4,
modifier = Modifier.fillMaxWidth(),
)
// model + language
Row(horizontalArrangement = Arrangement.spacedBy(12.dp)) {
ExposedDropdownMenuBox(
@@ -62,6 +62,7 @@ data class SessionConfig(
val liveModel: String?,
val liveIntervalSec: Int = 8,
val storageUrl: String? = null,
val agenda: List<String> = emptyList(),
)
/** Recording state published to the UI (collect RecorderService.state). */
@@ -326,7 +327,7 @@ class RecorderService : Service() {
durationMs = durationMs,
language = cfg.language ?: "",
device = "Android",
agenda = emptyList(),
agenda = cfg.agenda,
)
}
Log.i(TAG, "upload done: $id")
@@ -17,6 +17,19 @@ object StorageClient {
val language: String,
val device: String,
val files: List<String>,
val agenda: List<String> = emptyList(),
val summary: String? = null, // null | "pending" | "done" | "error: ..."
val agendaStatus: String? = null,
val agendaResults: List<AgendaResult>? = null,
)
/** Agenda coverage as judged by the LLM. */
data class AgendaResult(
val item: String,
val covered: Boolean,
val time: String,
val evidence: String,
val error: String? = null,
)
/** Lists recordings, newest first. */
@@ -34,25 +47,62 @@ object StorageClient {
files = o.optJSONArray("files")?.let { f ->
(0 until f.length()).map { j -> f.getString(j) }
} ?: emptyList(),
agenda = o.optJSONArray("agenda")?.let { a ->
(0 until a.length()).map { j -> a.getString(j) }
} ?: emptyList(),
summary = if (o.isNull("summary")) null else o.optString("summary"),
agendaStatus = if (o.isNull("agenda_status")) null else o.optString("agenda_status"),
agendaResults = o.optJSONArray("agenda_results")?.let { a ->
(0 until a.length()).mapNotNull { j ->
val r = a.optJSONObject(j) ?: return@mapNotNull null
AgendaResult(
item = r.optString("item", ""),
covered = r.optBoolean("covered", false),
time = r.optString("time", ""),
evidence = r.optString("evidence", ""),
error = r.optString("error", "").ifBlank { null },
)
}
},
)
}
}
/** Fetches a single recording's metadata (for polling summaries). */
fun fetchMeta(baseUrl: String, id: String): RecordingMeta =
list(baseUrl).first { it.id == id }
/** Downloads one of the recording's files (e.g. meeting.json). */
fun fetchFile(baseUrl: String, id: String, name: String): ByteArray =
get(url(baseUrl, "/api/recordings/$id/files/$name"))
/** Deletes a recording (204 expected). */
fun delete(baseUrl: String, id: String) {
val conn = (URL(url(baseUrl, "/api/recordings/$id"))
.openConnection() as HttpURLConnection).apply {
requestMethod = "DELETE"
post(baseUrl, "/api/recordings/$id", method = "DELETE")
}
/** Asks the server to (re)generate the summary via Ollama. */
fun triggerSummary(baseUrl: String, id: String) {
post(baseUrl, "/api/recordings/$id/summary")
}
/** Asks the server to (re)check agenda coverage via Ollama. */
fun triggerAgenda(baseUrl: String, id: String) {
post(baseUrl, "/api/recordings/$id/agenda")
}
private fun post(baseUrl: String, path: String, method: String = "POST") {
val conn = (URL(url(baseUrl, path)).openConnection() as HttpURLConnection).apply {
requestMethod = method
connectTimeout = 15_000
readTimeout = 30_000
readTimeout = 60_000
}
try {
if (conn.responseCode !in 200..299) {
throw IOException("server returned ${conn.responseCode}")
val code = conn.responseCode
if (code !in 200..299) {
val resp = conn.errorStream?.use { it.readBytes() } ?: ByteArray(0)
throw IOException("server returned $code: " +
String(resp, 0, minOf(resp.size, 200)))
}
} finally {
conn.disconnect()
+180 -1
View File
@@ -14,10 +14,12 @@ Storage layout (no database):
"""
import json
import os
import re
import shutil
import threading
import time
import urllib.request
import uuid
from datetime import datetime, timezone
from pathlib import Path
@@ -33,6 +35,14 @@ INDEX_PATH = DATA_DIR / "index.json"
ALLOWED_FILES = ("wav", "txt", "srt", "json")
RID_RE = re.compile(r"^[A-Za-z0-9-]+$")
OLLAMA_URL = os.environ.get("OLLAMA_URL", "").rstrip("/")
OLLAMA_MODEL = os.environ.get("OLLAMA_MODEL", "gemma4:12b")
OLLAMA_NUM_CTX = int(os.environ.get("OLLAMA_NUM_CTX", "32768"))
OLLAMA_TIMEOUT = int(os.environ.get("OLLAMA_TIMEOUT", "600"))
# rough token guard: ~4 chars per token; 90k chars fit 32k context
TRANSCRIPT_MAX_CHARS = 90_000
app = FastAPI(title="meetrec-server", version="0.1.0")
_lock = threading.Lock()
@@ -56,6 +66,143 @@ def _find(items: list, rid: str) -> Optional[dict]:
return next((it for it in items if it["id"] == rid), None)
def _update_item(rid: str, **fields) -> Optional[dict]:
"""Thread-safe index field update for one recording."""
with _lock:
items = _load_index()
item = _find(items, rid)
if item is None:
return None
item.update(fields)
_save_index(items)
return dict(item)
# ------------------------------------------------------------------ Ollama
def _ollama_chat(prompt: str) -> str:
body = json.dumps({
"model": OLLAMA_MODEL,
"messages": [{"role": "user", "content": prompt}],
"stream": False,
"options": {"num_ctx": OLLAMA_NUM_CTX, "temperature": 0.2},
}).encode()
req = urllib.request.Request(
OLLAMA_URL + "/api/chat", data=body,
headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(req, timeout=OLLAMA_TIMEOUT) as r:
data = json.load(r)
return data["message"]["content"].strip()
def _timestamped_transcript(rec_dir: Path) -> str:
"""'[h:]mm:ss text' lines from meeting.json, meeting.txt as fallback."""
js = rec_dir / "meeting.json"
if js.is_file():
try:
with js.open() as f:
segments = json.load(f).get("segments", [])
lines = []
for s in segments:
t = int(float(s.get("start", 0)))
stamp = f"{t // 3600}:{t % 3600 // 60:02d}:{t % 60:02d}" if t >= 3600 \
else f"{t // 60:02d}:{t % 60:02d}"
lines.append(f"[{stamp}] {s.get('text', '').strip()}")
if lines:
return "\n".join(lines)
except (json.JSONDecodeError, OSError, ValueError):
pass
txt = rec_dir / "meeting.txt"
return txt.read_text() if txt.is_file() else ""
_SUMMARY_PROMPT = """Du erstellst ein Meetingprotokoll auf Deutsch.\n\
Fasse das folgende Transkript in Markdown zusammen, mit den Abschnitten:
- **Thema** — das Thema des Meetings in einem Satz
- **Wichtigste Punkte** — Stichpunkte
- **Entscheidungen** — falls erkennbar
- **Aufgaben / To-dos** — wer macht was, falls erkennbar
Transkript:
{transcript}
Antworte nur mit der Zusammenfassung, ohne Vorwort.
"""
_AGENDA_PROMPT = """Prüfe für jeden Agendapunkt, ob er im folgenden Transkript\n\
besprochen wurde.
Antworte AUSSCHLIESSLICH mit JSON — kein Markdown, keine Erklärung:
[{{"item": "<Agendapunkt>", "covered": true oder false, "time": "<mm:ss des ersten Belegs, sonst leer>", "evidence": "<kurzes Zitat, maximal 15 Wörter, sonst leer>"}}]
Agenda:
{agenda}
Transkript:
{transcript}
"""
def _generate_summary(rid: str) -> None:
rec_dir = REC_DIR / rid
try:
transcript = _timestamped_transcript(rec_dir)
if not transcript.strip():
raise RuntimeError("no transcript available")
prompt = _SUMMARY_PROMPT.format(transcript=transcript[:TRANSCRIPT_MAX_CHARS])
(rec_dir / "summary.md").write_text(_ollama_chat(prompt), encoding="utf-8")
_update_item(rid, summary="done")
except Exception as e: # noqa: BLE001 — report in index, never crash the thread
_update_item(rid, summary=f"error: {str(e)[:120]}")
def _check_agenda(rid: str) -> None:
item = _update_item(rid) # read-only pass
items = item.get("agenda") if item else None
if not items:
return
rec_dir = REC_DIR / rid
try:
transcript = _timestamped_transcript(rec_dir)
if not transcript.strip():
raise RuntimeError("no transcript available")
agenda = "\n".join(f"- {a}" for a in items)
raw = _ollama_chat(
_AGENDA_PROMPT.format(agenda=agenda,
transcript=transcript[:TRANSCRIPT_MAX_CHARS]),
)
start, end = raw.find("["), raw.rfind("]")
if start < 0 or end <= start:
raise RuntimeError("model returned no JSON")
results = json.loads(raw[start:end + 1])
if not isinstance(results, list):
raise RuntimeError("model returned unexpected JSON")
clean = [
{
"item": str(r.get("item", ""))[:300],
"covered": bool(r.get("covered", False)),
"time": str(r.get("time", ""))[:20],
"evidence": str(r.get("evidence", ""))[:200],
}
for r in results if isinstance(r, dict)
]
(rec_dir / "agenda.json").write_text(
json.dumps(clean, ensure_ascii=False, indent=2), encoding="utf-8")
_update_item(rid, agenda_results=clean, agenda_status="done")
except Exception as e: # noqa: BLE001
_update_item(rid, agenda_results=[{"error": str(e)[:120]}],
agenda_status=f"error: {str(e)[:120]}")
def _postprocess(rid: str) -> None:
"""Background pipeline after an upload: summary, then agenda check."""
_generate_summary(rid)
_check_agenda(rid)
@app.get("/api/health")
def health() -> dict:
with _lock:
@@ -115,14 +262,18 @@ async def upload(
"device": device,
"agenda": agenda_items,
"files": sorted(stored),
"summary": None, # set in M3c
"summary": None, # none | pending | done | error: ...
"agenda_results": None,
"agenda_status": None,
"uploaded_at": time.time(),
}
with _lock:
items = _load_index()
items.append(item)
_save_index(items)
if OLLAMA_URL:
threading.Thread(target=_postprocess, args=(rid,), daemon=True).start()
return {"id": rid, "files": sorted(stored)}
@@ -151,6 +302,34 @@ def get_file(rid: str, name: str) -> FileResponse:
return FileResponse(path)
@app.post("/api/recordings/{rid}/summary", status_code=202)
def trigger_summary(rid: str, force: bool = False) -> dict:
with _lock:
item = _find(_load_index(), rid)
if item is None:
raise HTTPException(404, "unknown recording id")
if item.get("summary") == "pending" and not force:
return {"status": "pending"}
_update_item(rid, summary="pending")
threading.Thread(target=_generate_summary, args=(rid,), daemon=True).start()
return {"status": "started"}
@app.post("/api/recordings/{rid}/agenda", status_code=202)
def trigger_agenda(rid: str, force: bool = False) -> dict:
with _lock:
item = _find(_load_index(), rid)
if item is None:
raise HTTPException(404, "unknown recording id")
if not item.get("agenda"):
raise HTTPException(400, "recording has no agenda items")
if item.get("agenda_status") == "pending" and not force:
return {"status": "pending"}
_update_item(rid, agenda_status="pending")
threading.Thread(target=_check_agenda, args=(rid,), daemon=True).start()
return {"status": "started"}
@app.delete("/api/recordings/{rid}", status_code=204)
def delete_recording(rid: str) -> None:
if not RID_RE.match(rid):