Private retrieval · grounded answers
Find the moment. Open the proof.
MumbleFlow searches the meetings stored on your Mac with both exact-word and meaning-aware retrieval. Ask a question, inspect the strongest matches, and jump directly to the transcript evidence instead of trusting an answer with no source.
Apple Silicon · macOS 26+ · English · Models download during setup
Product sync
Let’s make the first run feel a little simpler. What still needs to happen before we launch?
We’re moving the launch to Thursday afternoon so we can give onboarding one more review.
I’ll take the onboarding review. I can have the updated checklist ready by Wednesday.
Great. Let’s keep the launch note short and calm. Mention the extra time for onboarding.
One thing is still open: do we send the update to everyone, or start with the pilot group?
Topics
Launch timing
A little more time for a careful first impression.
Onboarding review
A final pass through the first-run experience.
Timeline
The team reviewed the remaining launch work and agreed to make room for one more onboarding check.
Decisions
Move the launch to Thursday afternoon.
Next steps
Maya will finish the onboarding checklist by Wednesday.
Open questions
Should the update go to everyone or the pilot group first?
What you missed
Last 5 minutesA calm launch update
Keep the note short and mention the extra time for onboarding.
Next steps
Maya will finish the checklist by Wednesday.
Dictation History
We’re moving the launch to Thursday afternoon to give onboarding one more review. Thanks for the extra care as we get this ready.
2 minutes ago1. Review the onboarding flow.
2. Update the launch checklist.
3. Share the final note with the team.
Can you help me think through the simplest version of this idea? I want the first step to feel obvious.
1 hour agoSearch everything
The launch is moving to Thursday afternoon, giving the team time for one more onboarding review.
Hybrid retrieval
Search for words and meaning
A meeting might say “move the launch” when you remember “schedule change.” Exact search alone can miss that relationship; semantic search alone can underweight a name or precise phrase. MumbleFlow runs both locally, then fuses their rankings.
Exact-word recall
A blind-token FTS5 index supports BM25 ranking without storing recoverable source words in the search table.
Meaning-aware recall
Apple NLEmbedding sentence vectors find related passages even when your question and the transcript use different wording.
Balanced ranking
Equal-weight reciprocal-rank fusion combines both result lists so exact terms and semantic context can reinforce each other.
Grounded answers
Only top source chunks reach the local language model, and every answer citation must resolve to a record that was actually retrieved.
Timestamp navigation
Open a meeting result at the supporting transcript moment and, while retained audio exists, play the corresponding passage.
Encrypted vectors
The memory-resident USearch index is serialized as an AES-GCM-encrypted derived file rather than a readable embedding archive.
Answer discipline
A useful answer still needs evidence
Generated answers are an interface over retrieved sources—not permission for the model to improvise. The answer path is deliberately narrower than an open-ended chat.
- 01
Retrieve
MumbleFlow ranks the most relevant stored transcript chunks for your question.
- 02
Generate locally
Apple Intelligence or installed Qwen receives only the highest-ranked source text, not your entire meeting library.
- 03
Validate
Every returned citation ID must match one of the supplied sources. Unknown IDs are rejected.
- 04
Open the source
Select a valid citation to reach the meeting and supporting timestamp. Unsupported questions return an honest refusal.
Without a semantic generation model, search results still work. You lose the synthesized answer, not access to your own matching transcript passages.
Local data structures
Readable to you, opaque at rest
Meeting text and metadata are encrypted field by field. Lexical tokens are normalized and HMACed before entering FTS5, while the vector archive is encrypted separately. Encryption keys remain in this device's Keychain.
This is not a claim that the entire SQLite container uses SQLCipher. It is a more precise claim: customer content fields are sealed, lexical index terms are opaque, and the serialized semantic index is encrypted. Read the complete privacy and retention explanation.
Availability
External sources are not enabled yet
Meetings now. Connected sources later.
MumbleFlow contains read-only OAuth and indexing infrastructure for Gmail, Google Calendar, and Slack. The current downloadable build has no production client IDs, and public Google and Slack review work is unfinished. Those services therefore cannot be connected today.
When connector availability changes, this page will be updated with the exact supported scopes, sync behavior, deletion controls, and release version. Until then, use MumbleFlow search for meetings captured inside the app.
Questions
Frequently asked questions
What can I search in the current MumbleFlow download?
You can search locally stored meeting transcript passages. Gmail, Google Calendar, and Slack connector code exists, but those external sources are not enabled in the current download while production credentials and provider reviews remain unfinished.
Does meeting search work without Apple Intelligence?
Yes. Ranked search results do not require a generative model. A generated natural-language answer requires Apple Intelligence or the optional locally installed Qwen model; otherwise you can still open the matching transcript results directly.
How does MumbleFlow avoid unsupported search answers?
The answer generator receives only highly ranked source chunks and must return valid source IDs. MumbleFlow validates those IDs before showing citations and answers “I couldn’t find that” when the available evidence does not support a response.
Does search send meeting text to a server?
No. Meeting indexing, lexical and semantic retrieval, and answer generation run on your Mac. Search indexes and customer content are protected locally; network access is not needed for searching meetings after setup.
Keep exploring
The rest of MumbleFlow
All features
A complete tour of dictation, Oat meeting notes, search, personalization, privacy, and controls.
AI dictation for Mac
Natural voice typing that removes fillers, catches corrections, formats lists, and inserts finished text.
Oat AI meeting notes
Reviewable, actionable meeting notes with live catch-up, speaker memory, timelines, and source citations.
Meeting transcription
Separate microphone and computer audio, readable speaker turns, timestamp playback, and grounded notes.
Privacy and security
A plain-language account of local processing, encryption, retention, network access, and deletion.
About MumbleFlow
Why MumbleFlow exists, how it is built, and the product principles behind it.
Looking for technical detail? Read the local speech-to-text guide.