Background Jobs — Inngest

MeetAI uses Inngest for event-driven, durable background processing — specifically for post-meeting transcript formatting and AI summarization.

Why Inngest?

Inngest provides durable functions — multi-step workflows where each step is individually retried and checkpointed. If step 3 fails, it retries from step 3, not from step 1.

This is critical for MeetAI because the summarization pipeline involves:

A database write (mark meeting as “processing”)
A database read + complex transformation (resolve speaker names, merge fragments)
An external AI API call (Gemini — can be slow or flaky)
A final database write (save summary)

Without durable steps, a Gemini timeout would force re-running the entire pipeline.

Pipeline Architecture

Step-by-step Breakdown

Step 1: Mark Processing


const meeting = await step.run("mark-processing", async () => {
  const [meeting] = await db.update(meetings)
    .set({
      status: "processing",
      startedAt: startedAt ? new Date(startedAt) : undefined,
      endedAt: endedAt ? new Date(endedAt) : undefined,
    })
    .where(eq(meetings.id, meetingId))
    .returning();
 
  meeting.transcript.sort((a, b) => a.timestamp - b.timestamp);
  return meeting;
});

Setting status = "processing" immediately prevents the dashboard from showing stale data. Users see a “Processing…” indicator instead of an empty summary.

Step 2: Format Transcript

This step performs smart merge of consecutive speaker fragments:


// Before merge:
[10:01] User: "I want to..."
[10:02] User: "know about APIs."

// After merge:
[10:01] User: "I want to... know about APIs."

It also resolves speaker IDs to human-readable names by batch-querying the user and agents tables in parallel:


const [users, agentRows] = await Promise.all([
  db.select({ id: userTable.id, name: userTable.name })
    .from(userTable)
    .where(inArray(userTable.id, speakerIds)),
  db.select({ id: agents.id, name: agents.name })
    .from(agents)
    .where(inArray(agents.id, speakerIds))
]);

Step 3: AI Summarization

Uses @inngest/agent-kit with Gemini 2.5 Flash. The system prompt enforces a structured markdown output:


### Overview
Detailed, engaging summary of the session...
 
### Notes
Thematic sections with timestamp ranges (e.g., 05:00 - 10:00)

Step 4: Save Summary


await step.run("save-summary", async () => {
  await db.update(meetings)
    .set({
      summary: aiResponse,
      transcript: meeting.transcript,  // original (un-merged) for accuracy
      status: "completed",
    })
    .where(eq(meetings.id, meetingId));
});

Subtle detail: The merged transcript is sent to Gemini for better summaries, but the original un-merged transcript is saved to the database for playback accuracy. This is an intentional dual-representation design.

Event Schema


// Event sent from webhook handler
{
  name: "livekit/room_finished",
  data: {
    meetingId: string,   // nanoid meeting ID
    startedAt: number,   // epoch ms
    endedAt: number,     // epoch ms
  }
}

Inngest Dev Server

During local development, Inngest provides a dev server with a visual dashboard:


npx inngest-cli@latest dev

This allows inspecting event payloads, step execution logs, and retry behavior without deploying to production.

Failure Modes & Recovery

Failure	Inngest Behavior	User Impact
Gemini API timeout	Step 3 retries (exponential backoff)	Summary delayed, not lost
Neon DB connection drop	Steps 1/2/4 retry individually	Transparent to user
Invalid transcript (empty)	Early return with `{ success: false }`	Meeting marked complete, no summary
Partial step completion	Resumes from last failed step	No duplicate processing