shopping-questions

name

shopping-questions

description

Answer the user's questions about their shopping — both what they've already bought and what they could buy — by querying the shopping.db store. Use this whenever the user wants to recall a past purchase ("what was that nice gochujang I bought?", "did I get a fancy soy sauce last year — was it Japanese or Korean?"), asks what / when / how much they've bought, wants recommendations or to explore a cuisine or ingredient ("I'm in the mood for Malay food, what do you suggest?"), or asks whether a shop stocks something or what a good substitute would be for an unstocked item ("they don't have Camargue rice — what's close?"). Reach for it even when the question is vague or chatty. NOT for onboarding or removing shops (that's the manage-shops skill) or refreshing a catalogue.

Answering shopping questions

Everything lives in the SQLite store shopping.db. You answer questions with a two-part move: queries.py retrieves, you judge. The retrieval is deliberately broad and noisy; the value you add is reading the results and reasoning over them.

There are two question shapes:

• Recall — search what they bought ("what was that X I got?").
• Discovery — search the catalogue, including substitution when the literal thing isn't stocked ("explore Filipino food", "what's like Camargue rice?").

The golden rule: retrieval is noisy — always read and discard

queries.py uses LIKE, which is blunt. Never trust a row's mere presence:

• "red rice" surfaces a £120 cooking pot; "camargue" alone matches chocolate; "adobo" matches Mexican products (a cross-cuisine homonym); gift sets and cookbooks list dozens of ingredients in their blurb.
• A term in the title means the product is that thing ("Bombom Korean Gochujang…"). A term only in the description is weak ("this marinade pairs well with soy sauce" is not a soy sauce).

Two levers cut the worst of it, both built into the helper:

• Ingredient filter (default): restricts the catalogue to product_type='Ingredients', dropping cookware / tableware / bundles / cookbooks. Pass --all only when the user genuinely wants those.
• Title-matches first: results are ordered so "the product is an X" beats "the description mentions X".

Even after these, you still read every candidate and bin the junk before answering.

The helper

# What the user bought (includes delisted items; gives reorder URL + last-ordered):
uv run python queries.py purchases "gochujang" [--since 2025-01-01] [--until 2026-01-01] [--shop souschef]

# The live catalogue (ingredients only unless --all):
uv run python queries.py catalogue "carnaroli" [--all] [--shop souschef]

# Every distinct thing they've bought — read this to infer their taste:
uv run python queries.py profile

The same functions are importable (search_purchases, search_catalogue, list_purchases) if you need to compose something custom.

Recipe 1 — Recall: "what was that X I bought?"

1. queries.py purchases "<keyword>".
2. Read and filter. Discard rows that only matched on a description mention.
3. Answer with the product, when it was last bought, and the reorder URL — and temper any "you still have it" implication by recency (see "Bought" is not "still has" below). If they asked an attribute ("Japanese or Korean?", "which brand?"), read the title/description and say which, with a word on why.
4. If nothing genuinely matches, say so and offer the near-misses or a wider search. Don't manufacture a memory.

Verified example: "that nice gochujang paste" → purchases "gochujang" → Bombom Korean Gochujang Hot Sauce, last bought 2026-03-31, with its link.

Recipe 2 — Time-bounded recall

Add --since / --until (UK sense: "last year" = the previous calendar year; today's date is in the system prompt). If the window is empty, widen it and say you widened.

Verified example: "that fancy soy sauce last year" → purchases "soy sauce" --since 2025-01-01 --until 2026-01-01. Reading the results shows they're marinades and hot sauces that merely mention soy sauce — no standalone soy sauce. The honest answer: "You didn't buy a dedicated soy sauce in 2025; the closest were these Korean sauces — want me to look further back?"

Recipe 3 — Taste-aware discovery & substitution

This is where culinary knowledge turns a thin catalogue into good suggestions.

1. Read their taste. queries.py profile, then infer the pattern — both the forms (pastes, ferments, chilli oils, marinades, special rice…) and the cuisines they lean on. Weight everything below toward it.
2. Search the literal thing. catalogue "<cuisine or ingredient>".
3. Assess coverage honestly, then go adjacent. If results are thin or absent, don't stop — use your knowledge to brainstorm specific adjacent terms and search each:
   • Ingredient: named varieties and classic substitutes — for "Camargue rice" search "red rice", "carnaroli", "wild rice", "black rice", not the proper noun (which matches noise).
   • Under-stocked cuisine: its flavour anchors and shared-pantry staples — for Filipino: "calamansi", "banana ketchup", "coconut vinegar", "fish sauce", "soy sauce", "coconut", "pandan". Neighbouring cuisines that share a pantry are fair game.
4. Present substitutes in this shape (it mirrors how the user likes hard-to-find-ingredient notes): the product, why it's close, and what it won't quite cover flavour/texture-wise. Be candid about genuine gaps — "Sous Chef doesn't really do Filipino; banana ketchup and calamansi vinegar are about it" is a better answer than a padded list. And when you note they already own a component for a dish, run it through the recency check below — they may have bought that gochujang months ago and finished it.

Verified examples:

• "They don't stock Camargue rice" — actually they do (Organic Camargue Red Rice, £5.90), found via catalogue "red rice" not "camargue". If they didn't: Carnaroli (similar chew, less earthy/nutty), Chinese Black Rice (nuttier, stickier, bleeds colour), Red Wild Rice (chewier, grassier).
• "Explore Filipino" — honest: the catalogue is thin. Real finds are Baron Banana Ketchup (a genuine Filipino staple) and Calamansi Vinegar (the citrus acid behind so much Filipino food — sharper, less floral than fresh calamansi). Then, leaning on their Korean/SE-Asian taste, point at the soy–vinegar–garlic adobo backbone they can build from stocked soy sauces and vinegars, noting what's missing (no fresh calamansi, no bagoong shrimp paste).

After a discovery investigation, offer to save the findings

A Recipe 3 run usually ends in a sizeable, link-heavy spread of recommendations — worth keeping, not just scrolling past once. So once you've given more than a handful of suggestions, offer to write the full findings to a sensibly-named markdown file in the output directory (e.g. output/souschef-se-asia-ideas.md). Save the whole thing — the colour and the reasoning (why each suits their taste, the honest gaps, the overlap flags), not a bare list; that context is the point of the exercise. Format the links per Reorder links, prices and availability below (inline markdown urls are fine in the external file). Offer it; don't write the file unprompted.

"Bought" is not "still has" — temper by recency and shelf-life

Purchase history tells you what the user bought and when — never what's in the cupboard right now. Don't state possession as fact ("you've already got X"); they may have run through it long ago. queries.py purchases gives last_ordered and a friendly "~N months ago" — lead with that, and slide your confidence by how long that kind of thing lasts:

• Long-life pantry (dried rice/noodles, sealed spices, vinegar, soy, sugar, honey): fine to assume they probably still have it for many months.
• Perishable / short-life (fresh kimchi, pickled veg, opened ferments, fresh produce, anything chilled): assume it's likely gone after a few months.
• Non-consumable (cookware, tableware, books): they still have it — no rebuy.

Always include the date and the reorder link so they can judge for themselves — your shelf-life guess can be wrong (some people burn through a sauce in a week). Rough phrasing ladder, slid by the product's longevity:

• recent, or long-life → "you picked up X in (~N weeks ago), so you've probably still got some — if you want to top up."
• a good while ago, and consumable → "you last bought X back in <month/year> (~N months ago), so you've likely worked through it — worth re-adding: ."
• non-consumable, any age → "you've already got X (bought )."

Be useful and honest: surface the date, make a sensible guess at whether it's run out, and never assert the cupboard's contents as fact.

Reorder links, prices and availability

Every matched product carries a url and current price. Include the link when recall is about rebuying. If useful, compare what they paid (the purchase's unit_price) to the current price and flag rises, or flag if a past purchase is now delisted.

Write URLs in full, on their own line — never as an inline [text](url) link. The user's terminal hides inline links: it shows only the link text and swallows the actual URL, so a hidden link is useless to them. This applies everywhere you hand over a link — in chat and in any file you write. (Note the prices/links come from the catalogue snapshot in shopping.db; if it's gone stale a refresh may be due.)

When the answer is "no" or "not sure"

Empty results, description-only matches, and real catalogue gaps are legitimate answers. Offer a wider search or an adjacent suggestion — never invent a purchase or a stock item.

Files

• queries.py — retrieval helpers + CLI (this skill's workhorse).
• db.py — connection/schema for shopping.db.
• Onboarding or removing a shop is the separate manage-shops skill.