Asking Your Library Real Questions
The five questions worth asking, how to scope them by source and date, and how to read a cited answer without trusting it blindly.
Before you start
- second-brain/01-connecting-your-first-mcp-client
What you'll learn
- Ask questions that use meaning, not just keywords
- Scope a query by source type and date range
- Read citations as evidence rather than decoration
- Notice when keyword results are masking a broken semantic index
Try it yourself
Open this in Glassy and follow along interactively.
A search box answers “where is the thing I typed.” A corpus answers “what do I know about this, and where did I learn it.” The difference is mostly in how you ask.
The five questions
These are the shapes that pay off. Try one before reading the rest.
1. Recall with disagreement
“What did I save about sleep and memory, and where do the sources disagree?”
Disagreement is the interesting part. Any assistant can summarize one article; finding that two of your sources contradict each other is something only your corpus can tell you.
2. Decision archaeology
“Why did we pick Postgres? Show me the note where we decided.”
Grounded answers cite. If the note does not exist, the honest answer is “I do not have that” — which is still useful, because now you know the decision was never written down.
3. The shape you forgot
“I saved something about attention and long context — I cannot remember the title.”
This is the semantic lane doing its job. Exact words do not matter here; describe the idea.
4. Scoped sweep
“Across my vault files only, what do I have on onboarding flows?”
Scope it and the answer gets smaller and more trustworthy. The tools take a sources filter:
bookmarks, notes, documents, vault, voice.
5. Time-boxed catch-up
“What did I capture in the last two weeks that I have not mentioned since?”
This is the question that makes a read-later pile worth keeping. It is also the one most likely to surface an orphan — a note nothing links to. That is the next lesson.
Reading an answer like an adult
Three habits:
- Open one citation. If it does not say what the answer claims, the query was too vague — not the corpus, and not the model.
- Ask for the counter-case. “What do I have that argues against this?” is one extra sentence and it is where the value is.
- Prefer narrower scope. Five notes in the right source type beats five hundred across all of them.
When search lies to you
Keyword results can look healthy while the semantic lane is broken. In the Knowledge Base, check
diagnostics on a query that returned nothing — a degraded entry names the failing lane instead
of hiding it in an empty list. If keyword hits are fine and meaning-based results are missing, the
index needs a rebuild, not a new query.
What’s next?
Walking the knowledge graph — the questions that need links, not text: what points here, how are these two connected, and which notes are islands.