Movie recommendation chatbot on Qdrant and OpenAI

What it is

This n8n workflow is a movie recommendation chatbot. One part loads a CSV of movies into a Qdrant collection, with an OpenAI embedding for each description. The other part is a chat window, where an AI Agent turns a request into a liked and a disliked example and calls a tool. The tool asks Qdrant for the three closest movies and returns their names, years and descriptions. Twenty-five nodes do the work, beside two sticky notes.

Who it suits

Developers and data teams who want to see recommendation by example in n8n, and anyone who wants to adapt it to products, articles or other items with a short description. Skip it if you want picks based on viewing history. It compares descriptions only.

You need a little n8n: import a file, add two credentials, set two addresses and prepare a CSV of movies.

Where to use it

Any n8n that can reach OpenAI, your Qdrant and the CSV’s address will run it. The chat opens inside n8n, so the picks stay in your editor.

• n8n Cloud: import the file into a new workflow there
• Your own n8n server: the same file and steps
• Other items: load a CSV with the same three columns for books, products or songs
• Another collection name: change it in Qdrant Vector Store and in the two Qdrant addresses

Setup

Download saves a zip with workflow.json and Mindket’s setup notes in your downloads folder. Qdrant can be a cloud cluster or a server of your own.

1. In n8n, create a workflow, open its menu, choose Import from File and pick workflow.json.
2. Add an OpenAI credential and select it in Embeddings OpenAI, OpenAI Chat Model, Embedding Recommendation Request with Open AI and Embedding Anti-Recommendation Request with Open AI.
3. Add a Qdrant credential and select it in Qdrant Vector Store, Calling Qdrant Recommendation API and Retrieving Recommended Movies Meta Data.
4. In those two HTTP Request nodes, replace <YOUR_QDRANT_URL> with your Qdrant address, such as https://your-cluster.cloud.qdrant.io:6333.
5. In Download movies CSV, replace <YOUR_CSV_URL> with the address of your CSV.
6. Save the workflow, click Test workflow on the manual trigger and wait for it to finish.
7. Open the chat and ask for a recommendation.

How to use it

Open the chat and say what you want and what you do not, for example a light romantic comedy but nothing gory. The agent hands the tool a liked and a disliked description, then lists three movies, best first, without their scores. It keeps the last few messages as memory.

Example

A product manager types that she wants something like a warm romantic comedy, but not a gory horror film. The agent passes the first phrase as the liked example and the second as the disliked one.

The tool embeds both and asks Qdrant for the three movies that sit closest to the liked example and furthest from the disliked one. The agent replies with three films, their years and a line about each.

For newcomers and experts

If you are new. n8n draws a workflow as boxes joined by lines, and each box, a node, does one job. This file holds two flows. The first loads the movies once. The second is a chat whose agent calls a tool, and that tool is the lower group of boxes in this same workflow.

For experts. Call n8n Workflow Tool finds its own workflow with $workflow.id, so no id needs editing, and Execute Workflow Trigger receives the agent’s two strings under query. The Qdrant calls are plain HTTP Request nodes, since the recommend query is not a mode of the Qdrant node.

Good to know

The original needed fixes before it could work for anyone else, and it still needs your own CSV and Qdrant address.

• Models: OpenAI’s deprecations page, read on 12 November 2026, lists gpt-4o-mini as shutting down on 14 January 2027. Its replacement, gpt-6-luna, supports function calling in Chat Completions only with reasoning effort off, so Mindket’s copy uses gpt-5.6-terra for the agent. The page does not list the embedding model text-embedding-3-small
• Your own data: the original read a CSV from its author’s own repository and called its author’s own Qdrant cluster. Mindket’s copy asks for <YOUR_CSV_URL> and <YOUR_QDRANT_URL>
• Collection name: the original loaded a collection named imdb but searched imdb_1000_open_ai, so a fresh setup searched an empty or missing collection. Mindket’s copy uses movies for both
• Release year: the original read a field named release_year that the loader never wrote, so the year came back empty. Mindket’s copy reads movie_release_date
• Request bodies: the Qdrant bodies are built with JSON.stringify, so a vector cannot be joined into the text the wrong way, and the id list follows the number of results
• A dislike is needed: the tool asks for a liked and a disliked description, so a request that names no dislike leaves the agent to make one up
• Run the load once: loading twice stores every movie again under a new id, and a movie can then appear twice in the picks
• Memory: the agent’s window memory lives in n8n’s own memory, is cleared after an idle hour and does not suit queue mode
• Billing: each chat turn uses an agent call and two embeddings, and the load embeds one description per row, all billed by OpenAI
• Your data: your requests go to OpenAI, and the movie descriptions go to OpenAI and Qdrant
• Credentials: the JSON holds none, so add the OpenAI and Qdrant credentials yourself