Kitchen Table Talk

Worth knowing that the big shop slogans use the one word version deliberately and correctly. Everyday low prices means ordinary low prices, prices for everyday life, not prices that happen each day. Seeing it right in the wild is what finally fixed it for me.

Small-to-big, two columns, one join. chunk(id, doc_id, section_id, text, embedding) and section(id, text). Search chunks, group hits by section, pass sections to the model. It is fifteen minutes of work and it is the single highest-leverage change in most naive pipelines.

The thing to check now that you couldn't two years ago: how it handles a household where one person is enthusiastic and one is not. The good ones give the reluctant partner a read-only, low-notification view. That single feature has saved us roughly an argument a month.

Also normal to not need more than this. For a small internal tool, "does it still handle these 12 things people ask constantly" is a legitimate bar. Do not build an evaluation platform for a bot with 12k documents and 40 users.

Bought the LX four years ago mostly to kill a wobble. It has not moved a millimetre since. The thing nobody mentions: the spring tension is set with an Allen key and you have to actually set it for your monitor's weight. Out of the box it's tensioned for something heavier. Every person I've seen complain that their LX 'floats up on its own' skipped that step and blamed the arm.

Debounce the submit 300ms and move on. Most double submits are physical double clicks.

Something nobody mentions: pick the one that opens fastest. Sounds trivial, but a three-second cold start is enough friction to lose me on a Tuesday, and I'm not being dramatic about that.

That does dodge this specific error, since a route URL is stable across builds. But the stale tab is still running old JS against a new API and will 404 on chunks the moment it navigates, so you need the reload path regardless. You traded a loud failure for a quiet one.

Above the kettle. Of course it was above the kettle. Moving the whole shelf today.

The counterargument nobody likes: independent rating schemes do show real spread between helmets that all pass the same standard. Passing is a floor, not a ranking. Look up a SHARP score or equivalent for whatever you're considering before you buy.

Ouch, but fair.

This is the exact thing that broke my last two attempts and nobody has ever explained it that way.

Also make sure you know which direction you are measuring. In pgvector <=> is cosine distance, so smaller is better, and <#> is negative inner product. Half the "my similarity scores are backwards" posts are someone ordering by the wrong operator and getting a beautifully sorted list of the least relevant documents.

The genuinely useless rules I have dropped: rinsing mushrooms ruins them, searing seals in juices, and adding oil to pasta water stops it sticking. None of them survive five minutes of paying attention.

Just raise probes, that is what it is for.

Base64 the file into the action payload and you avoid the FormData handling entirely.

Politely disagreeing with the top answer on one point: foam is not the only thing you're buying. Uppers and outsole rubber on cheap shoes tend to go first, and I've had two budget pairs where the heel counter collapsed while the midsole was still fine. If your current pair has no creasing at the heel after eight months, ignore all of us.

Ferment the small ones whole, slice the big ones and put those in vinegar. Nothing keeps a fermented cucumber crunchy past midwinter, so stop trying.

Started at thirty six, finished at forty two, and the honest summary is that the age gap was the least significant thing about it. What was different in a good way: I had run projects before, so I could scope work, say no, and read a room in a meeting, and those three things are worth more in a PhD than raw cleverness. What was different in a bad way: the cohort bonds over being broke and free on Thursdays, and I had a mortgage and a school run, so I was never going to be at those pub nights and I had to build my social life around the department rather than inside it. The thing nobody warns you about is that you go from being the person who knows how everything works to being the person who does not know how anything works, and that is a genuine ego event around month four.

Requesting the SPD in writing also creates a record of when you asked, which matters if a window is later disputed.

I worked in a chain kitchen and the portioning rules for the expensive proteins were completely different. Everything else was a scoop, the expensive one was counted by piece and weighed at close, and waste on it was reported separately. That alone tells you where the margin risk sits.

Get one block rather than two, plus a strap, and put the rest toward a better mat if yours is thin. A slippery mat ruins more poses than a missing block.

Worth noting sqlx compile time checking is optional, you can use plain queries and get most of the way with a much gentler ramp.

Recall@50 was 0.71. Going through the misses, over half of them were queries containing a part number. Added a tsvector column and fused with RRF and it is 0.94. Then added the reranker and the top-5 got noticeably better on top of that.

This is the right framing. I paid for the wrong philosophy for a year and spent the whole time blaming the app.