Using It Well¶
The corpus has only two verbs: feed it and ask it. Both reward a little technique. This page is the practical guide to doing each well.
Phrasing good queries¶
The system answers by matching your question to topic pages, then reading them. So the whole game is: help it find the right pages, and tell it what kind of answer you want.
The three habits that matter most¶
1. Name the thing specifically. The more concrete the noun, the better the match.
- ✗ "Tell me about AI" — too broad; matches everything and nothing.
- ✓ "What do my sources say about mixture-of-experts?" — lands on a real page.
2. Use the words your sources use. The corpus tracks nicknames (GPT-4 / gpt4 / "GPT 4"), but exact terms hit hardest. If you saved things calling it "MoE," try "MoE."
3. Say what shape of answer you want. A definition? A comparison? A list of gotchas? The disagreement? Tell it.
Query patterns the system is built for¶
| You want… | Phrase it like… |
|---|---|
| Recall | "What do my sources say about context engineering?" |
| Compare / synthesize | "How do my notes compare RAG vs fine-tuning?" |
| Find the conflict | "Where do my sources disagree about agent memory?" |
| Connect ideas | "What's the relationship between parquet and columnar storage in what I've read?" |
| Scoped recall | "In my data-engineering pages, what are the gotchas with X?" |
Two phrasings that keep it honest¶
- Start with "what do my sources say…" — this keeps it in recall mode, and anything pulled from the web gets clearly labelled
[fresh — not yet in corpus]rather than blended in. - Treat it as a conversation. Ask a broad question, then drill: "now just that one source's take," "now the counterargument."
How to read the answer¶
- A claim with a citation = from your corpus, verifiable. Trust it.
[fresh — not yet in corpus]= a web gap-fill, less vetted; it gets properly ingested later.- "Coverage is thin" is a gift — it's telling you which domain to feed next.
- If it offers to save the synthesis as a page, say yes when the answer is a keeper.
Don't ask it¶
Real-time questions ("latest news on X"), pure general knowledge it never ingested, or anything where you need the exact original wording — for that, go read the source. The corpus is a map, not the territory.
Feeding a new domain well¶
A domain is just a topic folder. The key insight: domains emerge from clusters — you don't declare them, you feed them into existence.
1. Feed a cluster, not a single item. One article becomes a page somewhere; it doesn't make a domain. Feed three or more related sources so the topic has enough mass to stand on its own. Below that bar, the system folds it into a neighbour.
2. Curate at the source — your playlists and labels are the sorting. When you put videos in a themed playlist or label your email, you are pre-organising, and the system routes by that. Organise where you save, and the domain forms itself.
3. Quality over volume. A focused domain of twenty strong sources beats a bloated two hundred. Feed substantive things (a deep talk, a thorough article) over thin ones (a tweet, a listicle) — and leave the noise out. A playlist of hobby videos does not belong in a technical corpus.
4. Feed in a loop, don't dump. The rhythm that works:
feed a cluster → let the nightly ingest distill it → query it → see the thin spots → feed the thin spots → repeat
Don't drop fifty sources at once expecting magic. Feed a handful, let it compound, and let the corpus tell you what it's missing (that "coverage is thin" signal).
5. Don't force the name — let it route. If you're unsure something is a real area, just feed the sources and let the system place them. They'll either crystallise into a new domain (if distinct and numerous) or land as pages in an existing one. Both are correct; under-grown domains get folded back automatically.
A worked example: starting a robotics domain¶
- Days 1–3: add four or five solid robotics talks to a "Robotics" playlist; label two robotics newsletters.
- Overnight: the ingest distills them, and a
roboticsdomain forms (three distinct sources clears the bar). - Then ask it: "what do my sources say about robot learning?" — you instantly see what's covered and what's thin.
- Fill the gap: thin on hardware? Feed three hardware sources. It compounds.
Feeding anti-patterns¶
- One source, expecting a domain → it's just a page.
- Off-topic or low-quality sources → they dilute the domain.
- The same thing fed five ways → de-duplication collapses it; feed five different angles instead.
- Never querying → you never see the gaps, and the feed-then-query loop breaks.
The one-line version: Name the thing and say what answer you want; curate at the source, feed clusters not singles, then query early and often — the queries show you exactly where to feed next.