Grounded, cited answers
A grounded answer is one that draws only from your notes and cites the exact passage it used. This is the heart of Altretta's AI: whether you ask inside the app or from a connected client, the model queries your vault before it answers.
This is what a chat window can't give you. Paste your notes into a generic assistant and you get a fluent answer with no way to check it. Altretta hands you the answer and its receipts.
Ask inside Altretta
You don't need an external client to get grounded answers. Altretta has query-first AI built in:
- Ask panel — pose a question about your vault and get a cited answer in a dedicated result panel that shows the answer, its source anchors, and a per-passage relevance reading.
- Chat — a conversation with your vault, with a per-vault conversation list and durable history, so your threads persist and stay organized by vault.
Both run the same query-first retrieval as a connected external model — they pull the passages that bear on your question and instruct the model to answer only from them.
Each question is retrieved and prompted on its own: the thread you see is kept by the app, but earlier turns are not re-sent to the model. If a follow-up depends on something you said three messages ago, restate it.
How to read a citation
Every claim points back to a specific span of a specific file, written as source:lines:
[decisions/2026-03-db.md:12-19]
That means the answer came from lines 12–19 of decisions/2026-03-db.md. Open the note, read the lines, and confirm for yourself. No claim floats free of its source.
The groundedness verdict
Not every question has a clean answer in your notes — and Altretta says so instead of papering over it. Every answer carries a groundedness verdict:
| Verdict | What it means |
|---|---|
grounded | At least two sufficiently relevant passages were found and cited. One strong passage is never enough on its own. |
weak | Some evidence exists, but it's thin or partial — treat the answer with caution. |
ungrounded | No adequate support was found. The model is told not to fill the gap with guesses. |
grounded deliberately requires corroboration, not just a high score: a single matching passage, however close, is reported as weak. So a topic you've written about exactly once will honestly say "partial support" rather than claiming certainty.
When a connected AI asks over MCP and your vault has nothing to support the question, it is told the question is unanswerable and instructed to say it doesn't know rather than invent facts.
An honest "I don't have enough in your notes to answer that" is a feature, not a failure. It tells you where your knowledge has gaps — and keeps the AI from manufacturing decisions, quotes, or notes you never wrote.
Want that refusal to be strict? Turn on Require grounding in the Trust panel. With it on, a merely weak verdict is declined too, not just an empty one.
Why it's query-first (and cheaper)
Instead of stuffing your entire vault into the prompt, the model calls aingle_ground, which queries the semantic graph and returns just the passages relevant to your question — each with a signed provenance anchor proving what the note said — plus an instruction the model must follow: cite the sources, respect the verdict, don't guess.
The payoff is structural: the model receives a handful of passages rather than your whole vault, so a question costs a fraction of the input tokens that pasting everything would — and every answer stays traceable.
The honesty rules
Grounded answers hold the model to a small, strict contract:
- Cite everything. Every claim resolves to
source:lines. - Respect the groundedness verdict. Weak or missing evidence is surfaced, not hidden.
- Never invent. No fabricated notes, quotes, or decisions.
- Flag a stale index. A stale index is reported as its own distinct state — not quietly as "nothing found". Altretta tells you to rebuild the index rather than letting the AI claim your notes are empty.
If answers seem to miss recent notes, your index may be stale. Rebuild it so the AI sees your latest writing before you trust the next answer.
Next step
Want to see — and control — exactly what context the AI receives? → What your AI sees. Want to pick which model answers? → AI profiles.
Related
AI you can trust
Back to the AI section overview.
Provenance & proofs
The signed anchor behind every citation — inspect and verify it in the Proof Inspector.
The semantic graph
The graph the model queries to pull only the passages that bear on your question.
Ask questions about your code
The same grounded, cited answers over an ingested codebase.
MCP tools reference
aingle_ground and the other tools behind a grounded answer.
Glossary
Groundedness, verdict, provenance anchor — defined.