Quick overview
This workflow ingests a local PDF into Qdrant with Ollama embeddings, then supports hybrid retrieval by querying Qdrant with both dense vectors and BM25 sparse vectors from an n8n chat trigger.
How it works
- Starts manually to read a PDF from disk and extract its text content.
- Checks whether the Qdrant collection exists and creates it with a 768-dimension dense vector and a BM25-based sparse vector field if needed.
- Splits the extracted text into chunks, adds metadata, generates dense embeddings with Ollama (nomic-embed-text), and inserts the documents into the Qdrant vector store.
- Scrolls all stored points from Qdrant, builds a per-point BM25 sparse vector payload from each point’s content, and updates Qdrant vectors without overwriting existing fields.
- Triggers on an incoming chat message, generates an embedding for the query via Ollama’s embeddings HTTP API, and runs a Qdrant hybrid search that fuses dense and BM25 results using RRF.
Setup
- Configure Qdrant credentials (REST API and Vector Store) and ensure the workflow points to the correct Qdrant URL/collection name (default: collection "testing").
- Set up Ollama credentials and ensure the nomic-embed-text:latest model is available, and update the embeddings endpoint URL if your Ollama host is not
http://host.docker.internal:11434. - Place the source PDF on the n8n host and update the file path in the disk read step (default:
/tmp/n8n_Self_Hosted_Enterprise_Terms_and_Conditions.pdf). - Enable the chat trigger and use its webhook/chat entry point to send queries into the workflow.