Description
Overview
This recipe recommendation automation workflow provides a no-code integration pipeline designed for personalized meal suggestions based on HelloFresh’s current weekly menu. It leverages an orchestration pipeline combining web scraping, vector embeddings, and semantic search to extract, process, and recommend recipes dynamically each week using HTTP request and manual trigger nodes.
Key Benefits
- Automates extraction of up to 10 recipe courses from HelloFresh’s weekly menu via HTML and JSON parsing.
- Transforms recipe data into structured documents for efficient vectorization and semantic similarity search.
- Utilizes vector embeddings and Qdrant Recommend API for personalized, context-aware recipe recommendations.
- Integrates an AI chat agent enabling interactive, preference-driven recipe queries and suggestions.
Product Overview
The workflow initiates manually using a manual trigger node and sends an HTTP GET request to fetch the HelloFresh weekly menu page. It dynamically constructs the URL using the current year and week number to ensure up-to-date data retrieval. The HTML response is parsed to extract server-side rendered JSON data embedded in a script tag. This JSON is then used to isolate the top 10 available courses by slicing the array of recipes.
For each course, metadata—such as recipe name, cuisines, category, tags, nutrition facts, recipe URL, and unique identifiers—is extracted and assigned. Subsequently, the workflow performs an HTTP request to retrieve the full recipe webpage and uses CSS selectors to extract detailed recipe components including description, ingredients, utensils, and instructions.
These details are merged and formatted into unified structured documents capturing all relevant recipe information. The documents undergo recursive character text splitting to create manageable text chunks. Vector embeddings of these chunks are generated using the Mistral Cloud embedding model, facilitating semantic vectorization for recommendation queries.
The embeddings and metadata are inserted into a Qdrant vectorstore collection named “hello_fresh,” supporting efficient similarity and recommendation searches. The original recipe documents with metadata are concurrently persisted in a local SQLite database. This dual storage supports both vector-based search and full document retrieval.
An AI chat agent powered by a Mistral Cloud chat model is integrated, which accepts user input describing desired and undesired recipe attributes. The agent leverages the Qdrant Recommend API to perform vector similarity queries that incorporate positive and negative preferences, filtered by the current week, and grouped by recipe ID for unique recommendations. A rate limit wait node manages external API call pacing to maintain stability.
Features and Outcomes
Core Automation
This automation workflow processes HelloFresh’s weekly menu recipes through data extraction, metadata enrichment, and vector embedding to enable personalized recommendations. The core orchestration pipeline applies deterministic steps to transform raw HTML into semantically searchable recipe documents.
- Extracts and processes up to 10 recipes per weekly menu in a single pass.
- Combines HTTP request, HTML extraction, and code nodes for comprehensive data transformation.
- Implements synchronous and asynchronous node execution for reliable data flow.
Integrations and Intake
The workflow integrates HTTP request nodes for accessing HelloFresh’s website and external APIs. Authentication is handled via predefined credentials for Mistral Cloud and Qdrant API nodes. Input payloads consist of JSON data extracted from HTML script tags and user chat queries capturing preference descriptions.
- HTTP Request node fetches weekly menu with dynamic URL parameters (year and week).
- Mistral Cloud API provides vector embeddings and chat model with API key authentication.
- Qdrant API interaction uses token-based credentials to perform vector similarity and recommendation queries.
Outputs and Consumption
The workflow outputs structured recipe documents enriched with metadata and vector embeddings. Recommendations are delivered as JSON objects containing recipe IDs and metadata groups, which are then resolved to full recipe texts via SQLite queries. The AI chat agent formats responses suitable for conversational consumption.
- Recipe documents formatted as structured text with sections for ingredients, utensils, nutrition, and instructions.
- Qdrant Recommend API returns grouped recipe recommendations filtered by current week.
- Final output delivered synchronously to the AI agent for user interaction.
Workflow — End-to-End Execution
Step 1: Trigger
The workflow initiates manually via a manual trigger node (“When clicking ‘Test workflow'”). This allows explicit user control to start the extraction and recommendation cycle on demand.
Step 2: Processing
An HTTP Request node fetches the current week’s HelloFresh menu HTML page using dynamic URL parameters for year and week number. An HTML extraction node parses the embedded JSON script tag containing recipe data. The workflow performs basic presence checks and JSON parsing to extract the first 10 courses for further processing.
Step 3: Analysis
The workflow extracts detailed metadata and recipe content by sending HTTP requests to individual recipe URLs. HTML extraction nodes isolate descriptive fields, ingredients, utensils, and instructions using CSS selectors. Structured documents are prepared and split into chunks. Vector embeddings are computed using Mistral Cloud’s embedding API. The Qdrant Recommend API is then called with vectorized positive and negative user queries, applying filtering by the current week and grouping results to ensure unique recipe recommendations.
Step 4: Delivery
The recommendation results from Qdrant are processed to retrieve full recipe documents from a local SQLite database. These documents are assembled and passed to the AI chat agent, which generates a conversational response. The workflow delivers the final output synchronously through the chat interface triggered by webhook events.
Use Cases
Scenario 1
A user wants personalized dinner suggestions from the current HelloFresh menu but needs to avoid certain allergens. The orchestration pipeline processes weekly recipes and uses vector-based queries with positive and negative preferences to recommend suitable meals. The outcome is a filtered list of recipes matching their dietary needs in one response cycle.
Scenario 2
A meal planner requires quick access to the top recipes available this week for meal prep. By automating data extraction and vectorization, the workflow provides a semantic search interface via AI chat, returning relevant meal options instantly without manual browsing or filtering.
Scenario 3
A developer seeks to build a recommendation engine based on HelloFresh’s weekly menu data. This workflow demonstrates an end-to-end solution combining web scraping, document preparation, vector embedding, and AI-driven querying, ensuring scalable and maintainable recipe recommendations aligned with current offerings.
How to use
After importing this workflow into n8n, set up API credentials for Mistral Cloud and Qdrant as required. Confirm the existence of the Qdrant collection named “hello_fresh” with appropriate vector size and distance metric. Trigger the workflow manually to fetch and process the current week’s menu. Configure webhook access for the AI chat agent to enable interactive recipe queries. Users can provide positive and negative preferences to receive customized recipe recommendations. Expect structured textual recipe details returned via the AI agent in real time.
Comparison — Manual Process vs. Automation Workflow
| Attribute | Manual/Alternative | This Workflow |
|---|---|---|
| Steps required | Multiple manual steps: browsing, copying recipes, manual filtering | Single automated pipeline executing extraction, embedding, and recommendation |
| Consistency | Inconsistent due to human error and subjective filtering | Deterministic processing with structured extraction and vector search |
| Scalability | Limited by manual capacity and time constraints | Scalable with API-driven data retrieval and vectorstore indexing |
| Maintenance | Ongoing manual updates and re-collection needed weekly | Maintenance involves credential updates and API endpoint validation only |
Technical Specifications
| Environment | n8n automation platform with external API access |
|---|---|
| Tools / APIs | HelloFresh web pages, Mistral Cloud embedding and chat APIs, Qdrant vectorstore API, SQLite database |
| Execution Model | Combination of synchronous HTTP requests and asynchronous node execution |
| Input Formats | HTML pages, embedded JSON, user chat queries (JSON) |
| Output Formats | Structured text documents, JSON recommendation responses |
| Data Handling | Transient processing of HTML and JSON data; vector embeddings stored in Qdrant; full documents saved in SQLite |
| Known Constraints | Relies on external API availability and HelloFresh website structure consistency |
| Credentials | API keys for Mistral Cloud and Qdrant services |
Implementation Requirements
- Valid API credentials for Mistral Cloud embedding and chat services.
- Configured Qdrant vectorstore collection named “hello_fresh” supporting 1024-dimensional cosine vectors.
- Access to HelloFresh weekly menu website for HTTP requests and data extraction.
Configuration & Validation
- Verify manual trigger node activates the workflow and initiates the HTTP request to HelloFresh menu URL.
- Confirm JSON extraction node correctly parses the embedded JSON data and extracts recipe courses.
- Test AI chat agent integration with sample positive and negative preference inputs to ensure recommendations are returned.
Data Provenance
- Trigger node: manualTrigger initiates workflow on user command.
- Data extraction nodes: HTML and code nodes parse HelloFresh weekly menu and recipe pages.
- AI and vector nodes: Mistral Cloud embedding and chat models, Qdrant Recommend API provide vectorization and semantic search.
FAQ
How is the recipe recommendation automation workflow triggered?
The workflow is initiated manually using a manual trigger node, allowing controlled execution to fetch and process the current week’s recipes.
Which tools or models does the orchestration pipeline use?
This orchestration pipeline leverages Mistral Cloud’s embedding and chat models for vectorization and conversational AI, and Qdrant’s Recommend API for vector similarity searches.
What does the response look like for client consumption?
The response consists of structured recipe documents including metadata, ingredients, instructions, and nutrition, delivered via an AI chat agent in a conversational format.
Is any data persisted by the workflow?
The workflow stores vector embeddings and metadata in a Qdrant vectorstore and persists full recipe documents with metadata in a local SQLite database for retrieval.
How are errors handled in this integration flow?
The workflow relies on n8n’s default error handling; a wait node addresses API rate limits. No custom retry or backoff mechanisms are explicitly configured.
Conclusion
This recipe recommendation workflow automates the extraction, processing, and vector-based retrieval of HelloFresh’s weekly menu recipes, enabling personalized meal suggestions through a conversational AI agent. By combining semantic vector search via Qdrant and AI language understanding from Mistral Cloud, it delivers deterministic and context-aware recommendations filtered by user preferences. The workflow depends on the availability and consistency of external APIs and the HelloFresh website structure, which are critical constraints. Overall, it provides a robust foundation for scalable and maintainable recipe recommendation within the n8n automation environment.
Reviews
There are no reviews yet.