vault-ask

Ask your Obsidian vault with fail-closed citation validation.

A small, dependency-free question-answering tool over a folder of Markdown notes. It finds relevant notes, asks your LLM to answer only from those excerpts, and fails closed unless the output contains an exact [[wikilink]] copied from the selected source set.

$ vault-ask "what did I decide about the pricing model?"
Q: what did I decide about the pricing model?

Flat 49 EUR/month, no per-seat pricing, decided after the churn analysis.
[[Decisions/2026-Pricing|2026-Pricing]]

Notes consulted:
- [[Decisions/2026-Pricing|2026-Pricing]]
- [[Meetings/2026-01-pricing-review|2026-01-pricing-review]]

When the model emits the fixed refusal, returns no citation, or cites an unselected note, the validator returns the canonical refusal instead:

$ vault-ask "what is my bank account number?"
Q: what is my bank account number?

No note in the vault answers this question.

Why

A second brain is only useful if knowledge comes back out. Many "chat with your notes" tools require a vector database and an indexing pipeline. vault-ask instead uses no index or database: it runs ripgrep over your vault, ranks notes by term rarity, and hands focused excerpts to whatever LLM you already use.

How it works

1. Candidate search — ripgrep scans the whole vault in milliseconds.
2. IDF ranking — notes are scored by the rarity of the query terms they contain, so a rare, specific word (a project codename) outweighs a word that appears in hundreds of notes. No embeddings, no index, no warm-up.
3. Focused excerpts — only the headings and matching lines of the top notes are sent to the model (notes can be long).
4. Constrained prompt — the model is instructed to cite its claims, avoid outside knowledge, and emit a fixed refusal when evidence is insufficient.
5. Fail-closed output check — a non-refusal answer is accepted only when it contains at least one exact citation from the selected notes. Unknown or missing citations produce the fixed refusal.

This is a mechanical provenance check, not semantic entailment verification. A valid citation proves which selected note was referenced; it does not prove that the note supports every sentence in the answer.

In JSON output, the backward-compatible grounded field now means that exact selected-source citation validation passed. It does not mean semantic entailment was evaluated. Arbitrary paraphrases of refusal cannot be detected mechanically when they also contain a valid selected-source citation.

Nothing leaves your machine except what your own LLM command chooses to send.

Install

Requires Python 3.9+ and ripgrep (rg) on your PATH.

# pip (installs the `vault-ask` command)
pip install git+https://github.com/guillaumevele/vault-ask.git

Or run it as a single file, no install:

git clone https://github.com/guillaumevele/vault-ask.git
cd vault-ask
python3 vault_ask.py "your question"

No dependencies beyond the Python standard library and ripgrep.

Configure your LLM

vault-ask shells out to whatever LLM command you set in VAULT_ASK_LLM. The prompt is piped on stdin by default, or substituted for {prompt} if the command contains that placeholder.

# Local model via Ollama (prompt on stdin):
export VAULT_ASK_LLM='ollama run llama3.1'

# Simon Willison's `llm` CLI (any provider it supports):
export VAULT_ASK_LLM='llm -m gpt-4o-mini'

# A CLI that takes the prompt as an argument — use the {prompt} placeholder:
export VAULT_ASK_LLM='your-llm-cli --prompt {prompt}'

Point it at your vault once:

export OBSIDIAN_VAULT="$HOME/Obsidian/MyVault"

Usage

vault-ask "what did I decide about X?"
vault-ask --vault ~/notes "when is the contract renewal?"
vault-ask --limit 8 --json "summarize my pricing decisions"

No LLM? Use --sources-only to just rank the most relevant notes — a smart grep for your vault that needs no model at all:

vault-ask --sources-only "pricing model"
# Most relevant notes for: pricing model
# - [[Decisions/2026-pricing|2026-pricing]]
# - [[Meetings/2026-01-pricing-review|2026-01-pricing-review]]

What it's good at — and what it isn't

Good at: factual lookups where the words of your question point at a note — decisions, numbers, names, "what did I say about …" — with selected-source citation validation and explicit abstention states.

Not good at: abstract questions whose vocabulary differs from your notes (you ask "my funding strategy", the note says "tax credit"). Keyword retrieval can miss semantically related notes, and an allowed citation can still accompany an unsupported claim. Semantic recall and entailment verification are outside this zero-dependency tool's current scope.

Tests

python3 -m unittest discover -s tests

Related

voice-to-vault is the other half of the loop: it routes your voice captures into the Obsidian vault that vault-ask then answers questions about. One files your thoughts, the other brings them back.

License

MIT — see LICENSE.