Database

Database

The app uses SQLite via rusqlite (bundled — no system SQLite dependency). The database file notetaker.db is created in the working directory of the Tauri process.

---

Schema

meetings table

Column	Type	Notes
id	INTEGER PK AUTOINCREMENT
date	TEXT NOT NULL	toLocaleString() of recording end time
title	TEXT NOT NULL	Auto-generated: Meeting <date>
raw_transcript	TEXT NOT NULL	Plain text from WhisperX
markdown_summary	TEXT NOT NULL	Markdown/fallback text from the LangGraph workflow
speakers	TEXT	JSON array of speaker names (nullable)
tags	TEXT	JSON array of tag strings from LLM (nullable)
structured_summary	TEXT	Full structured JSON from LangGraph (nullable)
transcript_segments	TEXT	Versioned transcript evidence JSON (nullable)
schema_version	INTEGER	Structured transcript/summary schema version, defaults to 1 for old rows

meetings_fts virtual table (FTS5)

CREATE VIRTUAL TABLE meetings_fts
USING fts5(title, raw_transcript, markdown_summary, content='meetings', content_rowid='id');

The content= option links the FTS index to the meetings table without duplicating data. Three triggers (INSERT, UPDATE, DELETE) keep the index in sync automatically.

---

Migrations

On every startup, lib.rs attempts additive migrations for existing databases:

for sql in &[
    "ALTER TABLE meetings ADD COLUMN speakers TEXT",
    "ALTER TABLE meetings ADD COLUMN tags TEXT",
    "ALTER TABLE meetings ADD COLUMN structured_summary TEXT",
    "ALTER TABLE meetings ADD COLUMN transcript_segments TEXT",
    "ALTER TABLE meetings ADD COLUMN schema_version INTEGER DEFAULT 1",
] {
    if let Err(e) = conn.execute(sql, []) {
        if !e.to_string().contains("duplicate column name") {
            eprintln!("[DB Migration] {}: {}", sql, e);
        }
    }
}

SQLite does not support IF NOT EXISTS on ALTER TABLE. The duplicate column name error is treated as a no-op. Any other error is logged to stderr but does not crash the app.

---

IPC commands

Command	Parameters	Returns
save_meeting	date, title, raw_transcript, markdown_summary, speakers?, tags?, structured_summary?, transcript_segments?, schema_version?	inserted row id
get_meetings	—	Meeting[] ordered by id DESC
search_meetings	query: string	Meeting[] via FTS5 MATCH, falls back to get_meetings if query is empty

FTS5 search

search_meetings appends * to the query for prefix search — typing "intel" matches "intelligence". Special characters are stripped before the query is passed to SQLite.

let fts_query = format!("{}*", query.trim().replace('"', ""));

The sidebar search in the UI debounces the input by 300ms before calling search_meetings.

---

Structured summary JSON shape

{
  "schema_version": 2,
  "metadata": {"title": "Meeting Summary", "date": "2026-06-20", "tags": ["release"]},
  "tldr": "One sentence summary.",
  "participants": [
    {"name": "Alice", "role": "PM", "engagement_level": "high"}
  ],
  "metrics": [
    {"label": "Latency", "value": "68.428s", "trend": "flat"}
  ],
  "key_decisions": [
    {
      "decision": "Ship the Windows sidecar fix",
      "rationale": "The packaged app needs the PyInstaller binary",
      "owner": "Igor",
      "evidence_segment_ids": ["seg_000001"],
      "evidence_quote": "Ship the Windows sidecar fix",
      "confidence": 0.95,
      "inference": false
    }
  ],
  "action_items": [
    {
      "task": "Update the README and wiki",
      "assignee": "Igor",
      "due_date": null,
      "priority": "Medium",
      "status": "open",
      "evidence_segment_ids": ["seg_000002"],
      "evidence_quote": "Update the README and wiki",
      "confidence": 0.9,
      "inference": false
    }
  ],
  "summary_points": ["Important detail from the meeting."],
  "risks": [],
  "open_questions": [],
  "unresolved_topics": []
}

This object is serialized as a JSON string and stored in the structured_summary column. The frontend parses it on load.

If the LLM returns invalid summary JSON, the Python engine stores a transcript-derived fallback summary with summary_generation_warning instead of storing the old hard-failure TL;DR.