Ask

What grounded answers mean

Grounded means every substantive claim in an Ask reply should be traceable to material Overshow has already stored: OCR, transcripts, UI snapshots, or indexed document chunks. The system retrieves evidence, then synthesises it into readable prose. It is not a general-purpose chatbot that invents facts when your archive is silent.

That matters for knowledge work where verification beats fluency: you can open citations and confirm wording, timing, and source application before acting on an answer.

How Ask works

Ask runs in two broad phases:

Phase	What happens
Retrieval	The same semantic and keyword signals that power Search surface candidate passages, scoped by your filters and configuration.
Synthesis	On macOS 26+, a local language model turns retrieved snippets into an answer while staying tied to those sources.

Context assembly respects limits on how much material is included in each request, balancing coverage with focus. Nearby segments from the same source are merged so the model sees coherent passages rather than isolated fragments.

For anchored questions that name people, products, projects, identifiers, or domain terms, Ask also checks lexical-only search hits and cached answers for meaningful-anchor overlap before synthesis starts. Semantic candidates that pass the configured distance threshold, including semantic matches inside hybrid retrieval, can still ground paraphrased questions with little or no lexical overlap. Specific alpha-numeric identifiers, such as AB-123, AB 123, JIRA-123, TASK 123, or iOS 17.5, must match exactly even on semantic matches and profile-recall context. Low-signal wording such as reporting verbs, question words, and weekday/month/date/time words does not count as grounding. The overlap threshold scales with the query so short questions can still keep partial-topic evidence, while longer questions require stronger agreement. If every candidate is only a weak lexical neighbour, Ask returns insufficient context without invoking the local language model.

Ask versus Search

Dimension	Search	Ask
Primary output	Ranked list of hits	One narrative answer plus citations
Mental model	“Show me every place this appears”	“What did we conclude / decide / say?”
Exploration	You scan, compare, and open many cards	You read the synthesis, then drill into cited moments
Ranking	Keyword and semantic signals. tuned for lists	Retrieval supports a single best-effort context bundle
Verification	Implicit. you judge each hit	Explicit. citations point to evidence
When evidence is weak	Empty or short result sets	Decline, fallback labelling, or qualified text

When to use which

Use Search when you need exhaustive coverage, exact phrase hunting, or side-by-side comparison of many snippets. Use Ask when a short, sourced summary saves time. as long as you are willing to open citations for anything compliance-critical.

Citations and verification

Every useful answer includes inline numbered citations that tie statements back to stored content. Citation chips highlight the matching source row, and each row can expand in place to show the stored text before you jump to Sources or Timeline for deeper review.

Citation fields

Field	Meaning
Inline number	The answer claim that points to a cited source row
Matched text	The excerpt surfaced as evidence; read this before trusting paraphrase
Channel and time	Whether the evidence came from screen text, meeting audio, chat, or documents, and when it came
Expanded source	The full stored text, local indexing status, and links to Sources or Timeline

Fallback status and the full LLM path

Not every environment can run the full on-device stack at every moment. Ask surfaces fallback status so you can tell:

Situation	What you should expect
Full LLM answer	The on-device LLM produced a structured reply from retrieved context
Extractive fallback	Text may be closer to quoted retrieval than a rewritten narrative. still grounded, differently presented

The UI and diagnostics make this distinction explicit so you do not mistake a condensed extract for a free-form model essay.

When Ask declines. and what to do

Low confidence triggers honest decline rather than invention. Practical responses:

• Broaden time or remove an overly tight app_name filter.
• Rephrase toward terms that likely appear in captures.
• Switch to Search in hybrid or keyword mode to inspect raw hits.
• Confirm capture was running and permitted for that context.

Prompt configurations

Overshow ships four bundled prompt configurations so you can steer tone and structure without hand-authoring system prompts:

Configuration	Typical character
Grounded	Conservative tying to sources; default trust posture
Concise	Shorter answers when you want speed
Audio-focused	Emphasis on spoken evidence in the context bundle
Chain-of-thought	More explicit reasoning steps while remaining source-bound

Exact labels in the UI may vary slightly by version; the intent is the same. pick the shape of answer you need, not a different knowledge base.

Retrieval diagnostics and export

Ask exposes retrieval diagnostics for transparency and debugging:

Diagnostic area	Examples
Timings	Where latency went. retrieval vs model vs assembly
Backend	Which on-device path answered
Fallback trigger	Why extractive or shortened behaviour occurred
Prompt preview	What the model actually saw (redacted where appropriate)
Context stats	Count and character totals for included snippets

You can export diagnostics as JSON or CSV for tickets or internal analysis, and export Markdown context when you want to paste evidence into an external LLM or notebook. still under your control, not sent automatically by Overshow.

Privacy and on-device processing

Ask’s retrieval uses your local index; synthesis on supported Macs uses a local language model on-device. No query text is sent to external search providers as part of this feature. Any organisation-wide cloud policies elsewhere in the product remain separate. Ask is designed around local evidence first.

Quick Ask overlay

Quick Ask (typically Opt+Space) opens a lightweight overlay so you can pose a question without leaving the app in front of you. Workflow:

1. Invoke the shortcut from any workspace.
2. Type your question; retrieval runs against your archive.
3. Read the answer and citations; jump to Sources links where offered.
4. Dismiss the overlay to return to your previous context.

This complements the main Ask route in the desktop shell. use the overlay for interruptions, use the full view for longer review sessions.

Desktop UI integration

• Ask is the full-width review surface for grounded answers.
• Citations expand inline and can link into Sources for transcript and capture detail.
• Timeline remains the chronological place to inspect nearby moments around a cited source.

Tips for better answers

• Ask specific questions that name projects, people (as they appear in data), or time periods.
• Enable sensible context caps. too little context starves the model; too much dilutes focus.
• Pick prompt configs to match evidence type (audio-heavy meetings vs on-screen specs).
• If an answer feels vague, check diagnostics before blaming the model. retrieval may be thin.

Example questions

Situation	Example phrasing
Post-meeting recap	“What action items were agreed in yesterday’s planning call?”
Incident	“Did we mention rollback before the outage window closed?”
Spec drift	“What did we decide about the API rate limit in design review?”
Personal recall	“Which URL did I have open when discussing the invoice?”