feat: Optionally use a LLM for scraped ingredient parsing

backend/app/controller/recipe/recipe_controller.py

@@ -1,16 +1,15 @@
 import re
 
 from app.errors import NotFoundRequest, InvalidUsage
-from app.models.recipe import RecipeItems, RecipeTags
+from app.models import Household, RecipeItems, RecipeTags
 from flask import jsonify, Blueprint
 from flask_jwt_extended import jwt_required
 from app.helpers import validate_args, authorize_household
 from app.models import Recipe, Item, Tag
 from recipe_scrapers import scrape_me
 from recipe_scrapers._exceptions import SchemaOrgException, NoSchemaFoundInWildMode
-from ingredient_parser import parse_ingredient
-
 from app.service.file_has_access_or_download import file_has_access_or_download
+from app.service.ingredient_parsing import parseIngredients
 from .schemas import (
  SearchByNameRequest,
  AddRecipe,
@@ -194,6 +193,10 @@ def getAllFiltered(args, household_id):
 @authorize_household()
 @validate_args(ScrapeRecipe)
 def scrapeRecipe(args, household_id):
+ household = Household.find_by_id(household_id)
+ if not household:
+ raise NotFoundRequest()
+
  try:
  scraper = scrape_me(args["url"], wild_mode=True)
  except NoSchemaFoundInWildMode:
@@ -231,20 +234,16 @@ def scrapeRecipe(args, household_id):
  recipe.photo = scraper.image()
  recipe.source = args["url"]
  items = {}
- for ingredient in scraper.ingredients():
- parsed = parse_ingredient(ingredient)
- name = parsed.name.text if parsed.name else ingredient
- item = Item.find_by_name(household_id, name)
+ for ingredient in parseIngredients(scraper.ingredients(), household.language):
+ name = ingredient.name if ingredient.name else ingredient.originalText
+ item = Item.find_name_starts_with(household_id, name)
  if item:
- description = f"{parsed.amount[0].quantity if len(parsed.amount) > 0 else ''} {parsed.amount[0].unit if len(parsed.amount) > 0 else ''}"
- # description = description + (" " if description else "") + (parsed.comment.text if parsed.comment else "") # Usually cooking instructions
-
- items[ingredient] = item.obj_to_dict() | {
- "description": description,
+ items[ingredient.originalText] = item.obj_to_dict() | {
+ "description": ingredient.description,
  "optional": False,
  }
  else:
- items[ingredient] = None
+ items[ingredient.originalText] = None
  return jsonify(
  {
  "recipe": recipe.obj_to_dict(),

backend/app/models/item.py

@@ -158,6 +158,14 @@ def find_by_default_key(cls, household_id: int, default_key: str) -> Self:
  def find_by_id(cls, id) -> Self:
  return cls.query.filter(cls.id == id).first()
 
+ @classmethod
+ def find_name_starts_with(cls, household_id: int, starts_with: str) -> Self:
+ starts_with = starts_with.strip()
+ return cls.query.filter(
+ cls.household_id == household_id,
+ func.lower(cls.name).like(func.lower(starts_with) + "%"),
+ ).first()
+
  @classmethod
  def search_name(cls, name: str, household_id: int) -> list[Self]:
  item_count = 11

backend/app/service/ingredient_parsing.py

@@ -0,0 +1,97 @@
+from ingredient_parser import parse_ingredient
+from litellm import completion
+import json
+import os
+
+from app.config import SUPPORTED_LANGUAGES
+
+LLM_MODEL = os.getenv("LLM_MODEL")
+LLM_API_URL = os.getenv("LLM_API_URL")
+
+class IngredientParsingResult:
+ originalText: str = None
+ name: str = None
+ description: str = None
+
+ def __init__(self, original_text, name, description):
+ self.originalText = original_text
+ self.name = name
+ self.description = description
+
+ def __str__(self):
+ return f"{self.originalText} -> {self.name} ({self.description})"
+
+
+def parseNLP(ingredients: list[str]) -> list[IngredientParsingResult]:
+ def parseNLPSingle(ingredient):
+ parsed = parse_ingredient(ingredient)
+ name = parsed.name.text if parsed.name else None
+ description = f"{parsed.amount[0].quantity if len(parsed.amount) > 0 else ''} {parsed.amount[0].unit if len(parsed.amount) > 0 else ''}"
+ # description = description + (" " if description else "") + (parsed.comment.text if parsed.comment else "") # Usually cooking instructions
+ return IngredientParsingResult(ingredient, name, description)
+
+ return [parseNLPSingle(e) for e in ingredients]
+
+
+def parseLLM(
+ ingredients: list[str], targetLanguageCode: str = None
+) -> list[IngredientParsingResult]:
+ systemMessage = """
+You are a tool that returns only JSON in the form of [{"name": name, "description": description}, ...]. Split every string from the list into these two properties. You receive recipe ingredients and fill the name field with the singular name of the ingredient and everything else is the description. Translate the response into the specified language.
+
+For example in English:
+Given: ["300g of Rice", "2 Chocolates"] you return only:
+[{"name": "Rice", "description": "300g"}, {"name": "Chocolate", "description": "2"}]
+
+Return only JSON and nothing else.
+""" + (
+ f"Translate the response to {SUPPORTED_LANGUAGES[targetLanguageCode]}. Translate the JSON content to {SUPPORTED_LANGUAGES[targetLanguageCode]}. Your target language is {SUPPORTED_LANGUAGES[targetLanguageCode]}. Respond in {SUPPORTED_LANGUAGES[targetLanguageCode]} from the start."
+ if targetLanguageCode in SUPPORTED_LANGUAGES
+ else ""
+ )
+
+ response = completion(
+ model=LLM_MODEL,
+ api_base=LLM_API_URL,
+ # response_format={"type": "json_object"},
+ messages=[
+ {
+ "role": "system",
+ "content": systemMessage,
+ },
+ {
+ "role": "user",
+ "content": f"Translate the response to {SUPPORTED_LANGUAGES[targetLanguageCode]}. Translate the JSON content to {SUPPORTED_LANGUAGES[targetLanguageCode]}. Your target language is {SUPPORTED_LANGUAGES[targetLanguageCode]}. Respond in {SUPPORTED_LANGUAGES[targetLanguageCode]} from the start.",
+ },
+ {
+ "role": "user",
+ "content": json.dumps(ingredients),
+ },
+ ],
+ )
+
+ llmResponse = json.loads(response.choices[0].message.content)
+ if len(llmResponse) != len(ingredients):
+ return None
+ parsedIngredients = []
+ for i in range(len(llmResponse)):
+ parsedIngredients.append(
+ IngredientParsingResult(
+ ingredients[i], llmResponse[i]["name"], llmResponse[i]["description"]
+ )
+ )
+
+ return parsedIngredients
+
+
+def parseIngredients(
+ ingredients: list[str],
+ targetLanguageCode=None,
+) -> list[IngredientParsingResult]:
+ if LLM_MODEL:
+ try:
+ return parseLLM(ingredients, targetLanguageCode) or parseNLP(ingredients)
+ except Exception as e:
+ print("Error parsing ingredients:", e)
+
+ return parseNLP(ingredients)

backend/requirements.txt

@@ -1,6 +1,10 @@
+aiohappyeyeballs==2.3.4
+aiohttp==3.10.0
+aiosignal==1.3.1
 alembic==1.13.2
 amqp==5.2.0
 annotated-types==0.7.0
+anyio==4.4.0
 apispec==6.6.1
 appdirs==1.4.4
 APScheduler==3.10.4
@@ -26,7 +30,9 @@ cryptography==42.0.8
 cycler==0.12.1
 dbscan1d==0.2.2
 defusedxml==0.7.1
+distro==1.9.0
 extruct==0.17.0
+filelock==3.15.4
 flake8==7.1.0
 Flask==3.0.3
 Flask-APScheduler==1.13.1
@@ -39,23 +45,32 @@ Flask-SQLAlchemy==3.1.1
 flexcache==0.3
 flexparser==0.3.1
 fonttools==4.53.1
+frozenlist==1.4.1
+fsspec==2024.6.1
 future==1.0.0
 gevent==24.2.1
 greenlet==3.0.0rc3
 h11==0.14.0
 html5lib==1.1
 html_text==0.6.2
+httpcore==1.0.5
+httpx==0.27.0
+huggingface-hub==0.24.5
 idna==3.7
+importlib_metadata==8.2.0
 ingredient_parser_nlp==1.0.0
 iniconfig==2.0.0
 isodate==0.6.1
 itsdangerous==2.2.0
 Jinja2==3.1.4
 joblib==1.4.2
+jsonschema==4.23.0
+jsonschema-specifications==2023.12.1
 jstyleson==0.0.2
 kiwisolver==1.4.5
 kombu==5.3.7
 lark==1.1.9
+litellm==1.42.7
 lxml==5.2.2
 lxml_html_clean==0.1.1
 Mako==1.3.5
@@ -65,10 +80,12 @@ matplotlib==3.9.1
 mccabe==0.7.0
 mf2py==2.0.1
 mlxtend==0.23.1
+multidict==6.0.5
 mypy-extensions==1.0.0
 nltk==3.8.1
 numpy==2.0.0
 oic==1.7.0
+openai==1.37.1
 packaging==24.1
 pandas==2.2.2
 pathspec==0.12.1
@@ -101,21 +118,27 @@ python-engineio==4.9.1
 python-socketio==5.11.3
 pytz==2024.1
 pytz-deprecation-shim==0.1.0.post0
+PyYAML==6.0.1
 rdflib==7.0.0
 rdflib-jsonld==0.6.2
 recipe_scrapers==14.58.0
+referencing==0.35.1
 regex==2024.5.15
 requests==2.32.3
+rpds-py==0.19.1
 scikit-learn==1.5.1
 scipy==1.14.0
 setuptools==71.0.3
 setuptools-scm==8.1.0
 simple-websocket==1.0.0
 six==1.16.0
+sniffio==1.3.1
 soupsieve==2.5
 SQLAlchemy==2.0.31
 sqlite-icu==1.0
 threadpoolctl==3.5.0
+tiktoken==0.7.0
+tokenizers==0.19.1
 toml==0.10.2
 tomli==2.0.1
 tqdm==4.66.4
@@ -136,5 +159,7 @@ wcwidth==0.2.13
 webencodings==0.5.1
 Werkzeug==3.0.3
 wsproto==1.2.0
+yarl==1.9.4
+zipp==3.19.2
 zope.event==5.0
 zope.interface==6.4.post2

docs/docs/self-hosting/advanced.md

@@ -47,6 +47,7 @@ Environment variables for `tombursch/kitchenowl` and `tombursch/kitchenowl-backe
 | `APPLE_CLIENT_SECRET` | | |
 | `GOOGLE_CLIENT_ID` | | |
 | `GOOGLE_CLIENT_SECRET` | | |
+| `LLM_MODEL` | | Set a custom ingredient detection strategy for scraped recipes from the web. More at [Ingredient Parsing](./ingredient_parsing.md) |
 
 Additionally, to setting these environment variables you can also override the start command to scale the backend up.
 Add the following line or take a look at this exemplary [docker-compose.yml](https://github.com/TomBursch/kitchenowl/blob/main/docker-compose-postgres.yml) file:

docs/docs/self-hosting/ingredient_parsing.md

@@ -0,0 +1,25 @@
+# Ingredient Detection
+
+When scraping recipes from the web multiple strategies to map ingredients to household items are available.
+The scraping only returns a list of ingredients (like `3 cloves garlic`, `1/2 teaspoon salt`, `2 carrots chopped`) from which we need to extract the ingredient name and description. The names are then mapped to existing household items.
+
+The default method is to use a local [natural language processing (NLP) model](https://github.com/strangetom/ingredient-parser/) trained on English ingredients. To use this leave the `LLM_MODEL` environment variable empty.
+
+Alternatively, you can use a [Large Language Model (LLM)](https://github.com/BerriAI/litellm), multiple models are supported. Using a LLM uses more resources but can provide better results, especially for languages other than English.
+It can automatically translate the ingredient names to the household language for better item detection.
+
+### OpenAI
+
+To use OpenAi you need to set the following environment variables:
+
+- `LLM_MODEL`: The model name (e.g. `gpt-3.5-turbo`)
+- `OPENAI_API_KEY`: Your OpenAI API key
+
+### Ollama
+
+Ollama is a locally hosted LLM. You can also run this on a remote host (e.g. your desktop), and when the host is not reachable it will fall back to the NLP method. 
+To use Ollama you need to set the following environment variables:
+
+- `LLM_MODEL`: The model name prefixed with `ollama` (e.g. `ollama/llama3.1`)
+- `OPENAI_API_KEY`: Your OpenAI API key
+- `LLM_API_URL`: The URL of the Ollama server (e.g. `http://localhost:11434`)