Recipe recommendations with Qdrant and Mistral

Scrape web menu → Mistral embeddings → Qdrant vector store + Recommend API → SQLite backup → AI Agent chat recommendations

Who

Operators or developers who need to build recommendation features in n8n

You get

Build an AI chatbot that recommends this week's HelloFresh recipes based on user preferences and avoidances

You need

n8n · mistral

Steps

1. Fetch This Week's Menu

   Click "Test workflow" to trigger manually. The workflow requests https://www.hellofresh.co.uk/menus/{{ $now.year }}-W{{ $now.weekNumber }} via HTTP Request node, extracts server-side data from script#__NEXT_DATA__ using the HTML node, then parses the first 10 available courses with a Code node (JavaScript). Pages are large and may take time.

2. Create Recipe Documents

   For each course from the previous step, run in parallel: Set node extracts metadata (name, cuisines, category, tags, nutrition, url, id); HTTP Request node fetches the recipe detail page; HTML node extracts description, ingredients, utensils, and instructions. Then merge metadata and details by position with the Merge node, and format into a Markdown document with the Prepare Documents Set node, including title, website, ingredients, utensils, nutrition, instructions, plus week, cuisine, category, tag, id, name fields.

3. Vectorize and Store in Qdrant

   Send formatted documents to the Default Data Loader node with metadata (week, cuisine, category, tag, recipe_id) and enable Recursive Character Text Splitter. The Embeddings Mistral Cloud node converts text to vectors (requires mistralCloudApi credential). Finally write to the hello_fresh collection via Qdrant Vector Store node in insert mode (collection must be pre-created with Cosine distance, 1024 dimensions).

   Setup guidance · Source text。Qdrant server or console, create vector collection

   Copy prompt

   PUT collections/hello_fresh
   {
     "vectors": {
       "distance": "Cosine",
       "size": 1024
     }
   }

4. Save Original Documents to SQLite

   Use a Code node (Python) to connect to local SQLite database hello_fresh_1.db, create a recipes table (id TEXT PRIMARY KEY, name TEXT, data TEXT, cuisine TEXT, category TEXT, tag TEXT, week TEXT), and batch-insert the full documents. This provides the AI Agent with a source for retrieving complete recipe content later.

   Execution prompt · Source text。n8n Code node, Python mode

   Copy prompt

   import sqlite3
   con = sqlite3.connect("hello_fresh_1.db")
   
   cur = con.cursor()
   cur.execute("CREATE TABLE IF NOT EXISTS recipes (id TEXT PRIMARY KEY, name TEXT, data TEXT, cuisine TEXT, category TEXT, tag TEXT, week TEXT);")
   
   for item in _input.all():
     cur.execute('INSERT OR REPLACE INTO recipes VALUES(?,?,?,?,?,?,?)', (
       item.json.id,
       item.json.name,
       item.json.data,
       ','.join(item.json.cuisine),
       item.json.category,
       ','.join(item.json.tag),
       item.json.week
     ))
   
   con.commit()
   con.close()
   
   return [{ "affected_rows": len(_input.all()) }]

5. Configure AI Agent for Chat Recommendations

   Set up a Chat Trigger node as the chat entry point, connected to an AI Agent node. The AI Agent uses Mistral Cloud Chat Model (model mistral-large-2402, requires mistralCloudApi credential) and mounts the Qdrant Recommend API tool. This tool is implemented via a sub-workflow: first call the Mistral embeddings API to convert user positive/negative descriptions to vectors, then call Qdrant's /collections/hello_fresh/points/recommend/groups endpoint (average_vector strategy, filtered by metadata.week, grouped by metadata.recipe_id for deduplication), and finally query SQLite for the full recipes to return to the Agent.

   Execution prompt · Source text。AI Agent node systemMessage field

   Copy prompt

   =You are a recipe bot for the company, "Hello fresh". You will help the user choose which Hello Fresh recipe to choose from this week's menu. The current week is {{ $now.year }}-W{{ $now.weekNumber }}.
   Do not recommend any recipes other from the current week's menu. If there are no recipes to recommend, please ask the user to visit the website instead https://hellofresh.com.

   Execution prompt · Source text。Qdrant Recommend API tool (toolWorkflow node) description field

   Copy prompt

   Call this tool to get a recipe recommendation. Pass in the following params as a json object:
   * positives - a description of what the user wants to cook. This could be ingredients, flavours, utensils available, number of diners, type of meal etc.
   * negatives - a description of what the user wants to avoid in the recipe. This could be flavours to avoid, allergen considerations, conflicts with theme of meal etc.

   Execution prompt · Source text。Qdrant Recommend API tool (toolWorkflow node) inputSchema field

   Copy prompt

   {
   "type": "object",
   "properties": {
   	"positive": {
   		"type": "string",
   		"description": "a description of what the user wants to cook. This could be ingredients, flavours, utensils available, number of diners, type of meal etc."
   	},
      "negative": {
       "type": "string",
       "description": "a description of what the user wants to avoid in the recipe. This could be flavours to avoid, allergen considerations, conflicts with theme of meal etc."
     }
   }
   }

You are setting up a recipe recommendation workflow in n8n. Please confirm in order: 1) n8n instance is available with mistralCloudApi and qdrantApi credentials configured; 2) Qdrant has hello_fresh collection created (Cosine, 1024); 3) Understand the workflow has five stages: data scraping, document preparation, vector storage, SQLite backup, AI Agent; 4) AI Agent systemMessage and Recommend API tool parameters are configured per source; 5) HelloFresh site structure changes may break scraping and require manual verification.