Turn a static content list into a smart chat assistant that suggests the best items in seconds. Ideal for teams that want fast, relevant recommendations from a large catalog without manual sorting. Works well for media and ecommerce use cases.

Data is pulled from a file stored in GitHub, parsed, and enriched with OpenAI embeddings. The items are stored in a Qdrant vector database for fast similarity search. A chat entry point collects the user request. The agent builds positive and negative examples, creates query embeddings, and calls the Qdrant recommend API. Results are merged with item details, ranked by score, and returned as the top three picks with clear reasons. Memory keeps short context for follow up questions.

To set this up, you need access to a GitHub repo with your catalog file, an OpenAI API key, and a Qdrant cluster or cloud account. Index the catalog once with the manual trigger, then use the chat trigger for live requests. Expect faster discovery, less manual work, and scalable recommendations that adapt to any growing catalog.

• Chat trigger collects user requests and passes them to an AI agent with short term memory
• GitHub file fetch pulls a CSV catalog and parses item fields for indexing
• OpenAI embeddings convert item descriptions and user examples into vectors for search
• Token splitting and metadata loading keep rich item details linked to each vector
• Qdrant insert builds the collection and stores vectors for fast similarity search
• HTTP requests create embeddings for positive and negative user examples in real time
• Qdrant recommend API returns similar items and scores which are then merged with metadata
• Aggregation and field selection shape a clean top three response for the agent to present

• Reduce manual curation from hours to minutes by returning the top three matches automatically
• Automate up to 90 percent of repetitive selection work with a chat based request flow
• Improve match quality by using semantic search and negative signals to avoid bad fits
• Handle large catalogs as Qdrant scales to millions of items without slowing down
• Connect GitHub, OpenAI and Qdrant in one place for a simple recommendation stack

1. Import the template into n8n: Create a new workflow in n8n > Click the three dots menu > Select 'Import from File' > Choose the downloaded JSON file.
2. You'll need accounts with GitHub, OpenAI and Qdrant. See the Tools Required section above for links to create accounts with these services.
3. In the n8n credentials manager, create a GitHub credential. Use a personal access token with read access to the repo that holds your catalog file. Name the credential clearly so your team can find it later.
4. Open the GitHub node and select your new GitHub credential. Set the owner, repository, and file path to your CSV catalog. Save and test the node to confirm the file loads.
5. Create an OpenAI API Key in your OpenAI account. In n8n, create an OpenAI credential and paste the key. Select this credential in all OpenAI embedding and chat model nodes.
6. Set up Qdrant. In n8n, create a Qdrant credential with your Qdrant URL and API key. Open the Qdrant Vector Store node and choose this credential. Enter your collection name.
7. Check the data loader and token splitter nodes. Map the text field that contains your item description. Ensure key metadata fields like title and release date or category are included.
8. Open the Call n8n Workflow Tool node. Confirm the workflow reference points to your current workflow or the intended sub workflow. Update the workflow selection if the ID changed after import.
9. Open the two HTTP Request nodes that create embeddings for user examples. Confirm the OpenAI credential is selected and the request body uses the same embedding model as your index.
10. Open the Qdrant recommendation request node. Confirm the collection name matches the one you used during insert. Check any payload fields expected by your Qdrant setup.
11. Click the manual trigger and run once to index your catalog from GitHub into Qdrant. Verify vectors were inserted and no errors appear.
12. Copy the chat webhook URL from the chat trigger. Send a sample message with a positive and a negative description. Confirm you receive a top three list with scores and item details.
13. If results look off, check that the same embedding model is used for indexing and querying. Also confirm the CSV headers match the fields expected in the extract and loader nodes.