sterilize: recipe context + spell cleanup + defensive food.id preservation

Three improvements driven by Cobb's review of the fan-out output:

1. Recipe context. _parse_batch now accepts an optional recipe_context
   dict carrying recipe_name, recipe_description, and recipe_steps.
   preview_recipe builds the context from the Mealie recipe and passes
   it through. The Sonnet prompt has new USE RECIPE CONTEXT WHEN
   AMBIGUOUS rules: "1 cup flour" is ambiguous (AP / bread / cake);
   the cooking steps usually disambiguate ("knead until elastic" →
   bread flour, "sift with cocoa powder" + cake recipe → cake flour).
   Step text capped to 3000 chars so the user prompt stays modest;
   defaults to all-purpose flour when steps don't disambiguate.
   Brand/style hints in the description carry through too.

2. Spell + grammar cleanup. New SPELL/GRAMMAR CLEANUP rules in the
   prompt: silently fix typos in food and note ("tomatos" → "tomatoes",
   "chopped finly" → "chopped finely", "heavy cram" → "heavy cream").
   Normalize spacing. Critically: preserve EVERY semantic value —
   numeric quantities verbatim, every prep state, brand, color. When
   uncertain whether something is a typo or intentional ("yellow
   squash" is a real food, not a typo), keep it. Original strings
   stay in originalText for audit / rollback.

3. Defensive food.id preservation in apply_recipe. Three new safeguards
   protect against Sonnet hallucinations dropping live recipe data:

   a) If Sonnet returns a single all-null parsed item but the original
      Mealie row had a real food.id, pass the original through
      verbatim. (Sonnet probably parse-failed; never blank a real link.)

   b) When Sonnet returns a food name that we can't resolve in Mealie's
      catalog AND the original had a food.id, preserve the original
      link rather than emit food=null.

   c) When Sonnet explicitly returns food=null on the first child of
      an ingredient that originally had a food.id, treat that as a
      misread and preserve the original. Real section headers — where
      the original was ALREADY foodless — still pass through cleanly.

   Net effect: no apply path can drop a recipe's existing food
   reference. Sonnet can ADD food links (good), CHANGE them (good),
   or fail to parse (we keep what was there). It cannot remove them.

The is_new_food field also benefits from recipe context — Sonnet has
more evidence to set is_new_food=false (matched a known canonical)
when the steps confirm the ingredient identity.
This commit is contained in:
Kayos 2026-04-30 12:27:17 -07:00
parent 94c07ab156
commit 6bcf79e5dc

View file

@ -87,7 +87,37 @@ PARSE RULES (for the common 1→1 case):
- DO NOT wrap output in markdown fences.
- DO NOT include any prose before or after the JSON.
Input shape: {{"ingredients": ["str", "str", ...]}}
SPELL/GRAMMAR CLEANUP (no info loss):
- Silently fix obvious typos in food and note: "tomatos" "tomatoes",
"all-purpose flouur" "all-purpose flour", "chopped finly" "chopped finely",
"1 cup heavy cram" food: "heavy cream".
- Normalize spacing: "1 cup rice" quantity 1 / unit cup / food rice.
- Preserve EVERY semantic value: numeric quantities verbatim, every prep state,
brand, color, cooking method. If you're unsure whether something is a typo
or intentional (e.g. "yellow squash" is a real food, not a typo of "squash"),
keep it.
USE RECIPE CONTEXT WHEN AMBIGUOUS:
- The user prompt may include `recipe_name`, `recipe_description`, and
`recipe_steps` alongside the ingredients. When an ingredient is ambiguous
(e.g. just "1 cup flour" could be all-purpose, bread, cake, etc.), use
the cooking steps to disambiguate.
- Steps say "knead until smooth and elastic" bread flour
- Steps say "sift with cocoa powder" + recipe is a cake cake flour
- Steps don't disambiguate → default to all-purpose flour
- If the recipe description names a brand/style, use it.
- DO NOT add ingredients that aren't in the input list. Steps may mention
garnishes that weren't in the ingredient list — leave them out.
- When the steps confirm what an ingredient is, set is_new_food=false
(you have evidence it matches a known canonical) when possible.
Input shape:
{{"recipe_name": "...", "recipe_description": "...",
"recipe_steps": ["step text", ...],
"ingredients": ["str", "str", ...]}}
The recipe_name/description/steps fields are optional context they may
be empty or missing; ignore them in that case.
Output shape: {{"parses": [[{{...}}, {{...}}], [{{...}}], [{{...}}, {{...}}, {{...}}], ...]}}
The outer list MUST have the same length as the input list. Each inner list
MUST contain at least 1 item (use the all-null junk-fallback if needed).
@ -135,14 +165,19 @@ class Sterilizer:
"""Dry-run: parse all ingredients, return proposals without writing.
Each input ingredient produces one IngredientProposal whose
parsed_items list has length 1 (normal case) or N (fan-out)."""
parsed_items list has length 1 (normal case) or N (fan-out).
Recipe context (name + description + cooking steps) is bundled
into the Sonnet call so ambiguous ingredients ("1 cup flour")
can be disambiguated by what the recipe actually does with them."""
recipe = self.mealie.get_recipe(slug)
ingredients = recipe.get("recipeIngredient") or []
if not ingredients:
return {"slug": slug, "name": recipe.get("name"), "proposals": []}
strings = [_render_ingredient_for_parse(ing) for ing in ingredients]
parses_per_input = self._parse_batch(strings)
recipe_context = _build_recipe_context(recipe)
parses_per_input = self._parse_batch(strings, recipe_context=recipe_context)
proposals: list[IngredientProposal] = []
for i, (ing, items) in enumerate(zip(ingredients, parses_per_input)):
@ -205,6 +240,25 @@ class Sterilizer:
new_ingredients.append(dict(orig_ing))
continue
# Defensive: if Sonnet returned a single all-null item but the
# original ingredient had a real Mealie food link, this is
# almost certainly a hallucination/parse failure. Pass the
# original through unchanged rather than blank the row.
orig_food = (orig_ing.get("food") or {}) if isinstance(orig_ing.get("food"), dict) else {}
orig_food_id = orig_food.get("id")
orig_food_name = orig_food.get("name")
if (
len(items) == 1
and not (items[0].get("food") or "").strip()
and not (items[0].get("note") or "").strip()
and (items[0].get("quantity") in (None, ""))
and orig_food_id
):
# Sonnet returned all-null for a row that already had data.
# Preserve the original verbatim — never lose a recipe link.
new_ingredients.append(dict(orig_ing))
continue
for child_idx, parsed in enumerate(items):
if child_idx == 0:
# First child inherits id/refId/originalText from the
@ -236,11 +290,27 @@ class Sterilizer:
if food_id:
new_ing["food"] = {"id": food_id, "name": food_name}
new_ing["isFood"] = True
elif child_idx == 0 and orig_food_id:
# Sonnet wanted to set a different food but we
# couldn't resolve. Preserve the original link
# rather than blank — never lose recipe data.
new_ing["food"] = {"id": orig_food_id, "name": orig_food_name or ""}
new_ing["isFood"] = True
else:
new_ing["food"] = None
new_ing["isFood"] = False
else:
# Section header style — clear food, mark not-food
# Sonnet returned no food. Distinguish two cases:
if child_idx == 0 and orig_food_id:
# Original had a Mealie food link → defensive: KEEP
# it. Sonnet probably misread the input as a section
# header but the recipe author already linked it.
new_ing["food"] = {"id": orig_food_id, "name": orig_food_name or ""}
new_ing["isFood"] = True
else:
# True section header (original was already foodless)
# OR fan-out child (these are net-new rows; if Sonnet
# didn't pick a food for them, drop quietly)
new_ing["food"] = None
new_ing["isFood"] = False
@ -438,11 +508,26 @@ class Sterilizer:
# --- per-batch Sonnet call ---------------------------------------------
def _parse_batch(self, strings: list[str]) -> list[list[IngredientParse]]:
def _parse_batch(
self,
strings: list[str],
*,
recipe_context: dict | None = None,
) -> list[list[IngredientParse]]:
"""Returns list-of-lists matching the input length. Each inner list
is the parses derived from one input string (1 in normal case, N
for fan-out, never 0)."""
prompt = json.dumps({"ingredients": strings}, ensure_ascii=False)
for fan-out, never 0).
recipe_context is optional when provided, includes recipe_name,
recipe_description, and recipe_steps so Sonnet can disambiguate
unclear ingredients via the cooking steps."""
body: dict = {"ingredients": strings}
if recipe_context:
for k in ("recipe_name", "recipe_description", "recipe_steps"):
v = recipe_context.get(k)
if v:
body[k] = v
prompt = json.dumps(body, ensure_ascii=False)
try:
resp = self.forge.run(
prompt=prompt,
@ -506,6 +591,39 @@ class Sterilizer:
return index
def _build_recipe_context(recipe: dict) -> dict:
"""Pull name + description + cooking steps off a Mealie recipe in a
Sonnet-friendly shape. Capped to ~3000 chars total of step text so
the user prompt doesn't blow past Sonnet's reasonable input size."""
out: dict = {}
name = (recipe.get("name") or "").strip()
if name:
out["recipe_name"] = name[:200]
desc = (recipe.get("description") or "").strip()
if desc:
out["recipe_description"] = desc[:600]
instructions = recipe.get("recipeInstructions") or []
steps: list[str] = []
char_budget = 3000
for step in instructions:
if not isinstance(step, dict):
continue
text = (step.get("text") or "").strip()
if not text:
continue
if char_budget <= 0:
break
# Truncate individual step if needed
if len(text) > char_budget:
text = text[:char_budget] + ""
steps.append(text)
char_budget -= len(text)
if steps:
out["recipe_steps"] = steps
return out
def _render_ingredient_for_parse(ing: dict) -> str:
"""Best string representation of a Mealie ingredient for sending to Claude."""
if ing.get("originalText"):