From 4186ec6188b28662ba698d5825d7877ab5497f1e Mon Sep 17 00:00:00 2001 From: patrick-motsch Date: Mon, 16 Feb 2026 09:29:03 +0100 Subject: [PATCH] feat(teamsbot): stop command detection, session context for AI, context summarization for long sessions Co-authored-by: Cursor --- .../features/teamsbot/datamodelTeamsbot.py | 2 + .../features/teamsbot/routeFeatureTeamsbot.py | 1 + modules/features/teamsbot/service.py | 92 ++++++++++++++++++- 3 files changed, 94 insertions(+), 1 deletion(-) diff --git a/modules/features/teamsbot/datamodelTeamsbot.py b/modules/features/teamsbot/datamodelTeamsbot.py index b5523b5a..e4803097 100644 --- a/modules/features/teamsbot/datamodelTeamsbot.py +++ b/modules/features/teamsbot/datamodelTeamsbot.py @@ -78,6 +78,7 @@ class TeamsbotSession(BaseModel): startedByUserId: str = Field(description="User ID who started the session") bridgeSessionId: Optional[str] = Field(default=None, description="Session ID on the .NET Media Bridge") meetingChatId: Optional[str] = Field(default=None, description="Teams meeting chat ID for Graph API messages") + sessionContext: Optional[str] = Field(default=None, description="Custom context/knowledge provided by the user for this session") summary: Optional[str] = Field(default=None, description="AI-generated meeting summary (after session ends)") errorMessage: Optional[str] = Field(default=None, description="Error message if status is ERROR") transcriptSegmentCount: int = Field(default=0, description="Number of transcript segments in this session") @@ -200,6 +201,7 @@ class TeamsbotStartSessionRequest(BaseModel): backgroundImageUrl: Optional[str] = Field(default=None, description="Override background image for this session") connectionId: Optional[str] = Field(default=None, description="Microsoft connection ID for Graph API access") joinMode: Optional[TeamsbotJoinMode] = Field(default=None, description="How the bot joins: systemBot, anonymous, or userAccount. Defaults to systemBot if credentials configured, else anonymous.") + sessionContext: Optional[str] = Field(default=None, description="Custom context/knowledge to provide to the bot for this session (e.g. meeting agenda, documents, background info)") class TeamsbotSessionResponse(BaseModel): diff --git a/modules/features/teamsbot/routeFeatureTeamsbot.py b/modules/features/teamsbot/routeFeatureTeamsbot.py index 70aba37f..575094bf 100644 --- a/modules/features/teamsbot/routeFeatureTeamsbot.py +++ b/modules/features/teamsbot/routeFeatureTeamsbot.py @@ -178,6 +178,7 @@ async def startSession( meetingLink=cleanMeetingUrl, botName=body.botName or config.botName, backgroundImageUrl=body.backgroundImageUrl or config.backgroundImageUrl, + sessionContext=body.sessionContext, status=TeamsbotSessionStatus.PENDING, startedByUserId=str(context.user.id), ).model_dump() diff --git a/modules/features/teamsbot/service.py b/modules/features/teamsbot/service.py index 6cd1fa1d..8b246f1f 100644 --- a/modules/features/teamsbot/service.py +++ b/modules/features/teamsbot/service.py @@ -77,6 +77,8 @@ class TeamsbotService: # State self._lastAiCallTime: float = 0.0 self._contextBuffer: List[Dict[str, Any]] = [] + self._sessionContext: Optional[str] = None # User-provided background context + self._contextSummary: Optional[str] = None # AI-generated summary of long context # ========================================================================= # Session Lifecycle @@ -210,6 +212,13 @@ class TeamsbotService: interface = interfaceDb.getInterface(self.currentUser, self.mandateId, self.instanceId) voiceInterface = getVoiceInterface(self.currentUser, self.mandateId) + # Load session context (user-provided background knowledge) + session = interface.getSession(sessionId) + if session: + self._sessionContext = session.get("sessionContext") + if self._sessionContext: + logger.info(f"Session {sessionId}: Loaded session context ({len(self._sessionContext)} chars)") + logger.info(f"[WS-DEBUG] WebSocket handler started for session {sessionId}") try: @@ -323,6 +332,23 @@ class TeamsbotService: if not text: return + # Check for STOP command: " stop" or " STOP" + # This immediately stops the bot from speaking and clears the audio queue. + botNameLower = self.config.botName.lower() + textLower = text.lower() + if botNameLower in textLower and "stop" in textLower: + logger.info(f"Session {sessionId}: STOP command detected: [{speaker}] {text[:60]}") + if websocket: + try: + await websocket.send_text(json.dumps({ + "type": "stopAudio", + "sessionId": sessionId, + })) + except Exception as stopErr: + logger.warning(f"Failed to send stop command: {stopErr}") + # Don't trigger AI analysis for stop commands + return + # Filter out the bot's own speech from AI triggering. # The bot hears itself via captions — these should be stored in the # transcript for the record, but must NOT trigger AI analysis (feedback loop). @@ -351,6 +377,10 @@ class TeamsbotService: # Keep only last N segments maxSegments = self.config.contextWindowSegments if len(self._contextBuffer) > maxSegments: + # When buffer overflows, summarize the older half to preserve context + # without losing information. The summary replaces the old segments. + if not self._contextSummary and len(self._contextBuffer) > maxSegments * 1.5: + asyncio.create_task(self._summarizeContextBuffer(sessionId)) self._contextBuffer = self._contextBuffer[-maxSegments:] # Emit SSE event for live transcript @@ -473,7 +503,17 @@ class TeamsbotService: else: contextLines.append(f"[{speaker}]: {text}") - transcriptContext = f"BOT_NAME:{self.config.botName}\n" + "\n".join(contextLines) + # Include session context if provided by the user at session start + sessionContextStr = "" + if self._sessionContext: + sessionContextStr = f"\nSESSION_CONTEXT (background knowledge provided by the user):\n{self._sessionContext}\n" + + # Include summary of earlier conversation if available + summaryStr = "" + if self._contextSummary: + summaryStr = f"\nEARLIER_CONVERSATION_SUMMARY:\n{self._contextSummary}\n" + + transcriptContext = f"BOT_NAME:{self.config.botName}{sessionContextStr}{summaryStr}\nRECENT_TRANSCRIPT:\n" + "\n".join(contextLines) # Call SPEECH_TEAMS try: @@ -633,6 +673,56 @@ class TeamsbotService: logger.error(f"SPEECH_TEAMS analysis failed for session {sessionId}: {type(e).__name__}: {e}", exc_info=True) await _emitSessionEvent(sessionId, "error", {"message": f"AI analysis failed: {type(e).__name__}: {str(e)}"}) + # ========================================================================= + # Context Summarization (for long sessions) + # ========================================================================= + + async def _summarizeContextBuffer(self, sessionId: str): + """Summarize the older part of the context buffer to preserve information + without exceeding the context window. This runs in the background.""" + try: + if self._contextSummary: + return # Already summarized recently + + # Take the older half of the buffer for summarization + halfPoint = len(self._contextBuffer) // 2 + oldSegments = self._contextBuffer[:halfPoint] + + if len(oldSegments) < 10: + return # Not enough to summarize + + # Build text to summarize + lines = [] + for seg in oldSegments: + speaker = seg.get("speaker", "Unknown") + text = seg.get("text", "") + lines.append(f"[{speaker}]: {text}") + textToSummarize = "\n".join(lines) + + from modules.services.serviceAi.mainServiceAi import AiService + from modules.datamodels.datamodelAi import AiCallRequest, AiCallOptions, OperationTypeEnum, PriorityEnum + + serviceContext = _ServiceContext(self.currentUser, self.mandateId, self.instanceId) + aiService = AiService(serviceCenter=serviceContext) + await aiService.ensureAiObjectsInitialized() + + request = AiCallRequest( + prompt="Fasse das folgende Meeting-Transkript in 3-5 Saetzen zusammen. Nenne die wichtigsten Themen, Entscheidungen und offene Fragen. Antworte NUR mit der Zusammenfassung, keine Erklaerungen.", + context=textToSummarize, + options=AiCallOptions( + operationType=OperationTypeEnum.DATA_ANALYSE, + priority=PriorityEnum.SPEED, + ) + ) + + response = await aiService.callAi(request) + if response and response.errorCount == 0: + self._contextSummary = response.content.strip() + logger.info(f"Session {sessionId}: Context summarized ({len(oldSegments)} segments -> {len(self._contextSummary)} chars)") + + except Exception as e: + logger.warning(f"Context summarization failed for session {sessionId}: {e}") + # ========================================================================= # Meeting Summary # =========================================================================