Build a RAG system with automatic citations using Qdrant, Gemini & OpenAI

This workflow implements a Retrieval-Augmented Generation (RAG) system that:

• Stores vectorized documents in Qdrant,
• Retrieves relevant content based on user input,
• Generates AI answers using Google Gemini,
• Automatically cites the document sources (from Google Drive).

Workflow Steps

1. Create Qdrant Collection A REST API node creates a new collection in Qdrant with specified vector size (1536) and cosine similarity.

2. Load Files from Google Drive The workflow lists all files in a Google Drive folder, downloads them as plain text, and loops through each.

3. Text Preprocessing & Embedding

   • Documents are split into chunks (500 characters, with 50-character overlap).
   • Embeddings are created using OpenAI embeddings (text-embedding-3-small assumed).
   • Metadata (file name and ID) is attached to each chunk.

4. Store in Qdrant All vectors, along with metadata, are inserted into the Qdrant collection.

5. Chat Input & Retrieval

   • When a chat message is received, the question is embedded and matched against Qdrant.
   • Top 5 relevant document chunks are retrieved.
   • A Gemini model is used to generate the answer based on those sources.

6. Source Aggregation & Response

   • File IDs and names are deduplicated.

   • The AI response is combined with a list of cited documents (filenames).

   • Final output:

     AI Response
     
     Sources: ["Document1", "Document2"]
     

Main Advantages

• End-to-end Automation: From document ingestion to chat response generation, fully automated with no manual steps.
• Scalable Knowledge Base: Easy to expand by simply adding files to the Google Drive folder.
• Traceable Responses: Each answer includes its source files, increasing transparency and trustworthiness.
• Modular Design: Each step (embedding, storage, retrieval, response) is isolated and reusable.
• Multi-provider AI: Combines OpenAI (for embeddings) and Google Gemini (for chat), optimizing performance and flexibility.
• Secure & Customizable: Uses API credentials and configurable chunk size, collection name, etc.

How It Works

1. Document Processing & Vectorization

   • The workflow retrieves documents from a specified Google Drive folder.
   • Each file is downloaded, split into chunks (using a recursive text splitter), and converted into embeddings via OpenAI.
   • The embeddings, along with metadata (file ID and name), are stored in a Qdrant vector database under the collection negozio-emporio-verde.

2. Query Handling & Response Generation

   • When a user submits a chat message, the workflow:
     • Embeds the query using OpenAI.
     • Retrieves the top 5 relevant document chunks from Qdrant.
     • Uses Google Gemini to generate a response based on the retrieved context.
     • Aggregates and deduplicates the source file names from the retrieved chunks.
   • The final output includes both the AI-generated response and a list of source documents (e.g., Sources: ["FAQ.pdf", "Policy.txt"]).

Set Up Steps

1. Configure Qdrant Collection

   • Replace QDRANTURL and COLLECTION in the "Create collection" HTTP node to initialize the Qdrant collection with:
     • Vector size: 1536 (OpenAI embedding dimension).
     • Distance metric: Cosine.
   • Ensure the "Clear collection" node is configured to reset the collection if needed.

2. Google Drive & OpenAI Integration

   • Link the Google Drive node to the target folder (Test Negozio in this example).
   • Verify OpenAI and Google Gemini API credentials are correctly set in their respective nodes.

3. Metadata & Output Customization

   • Adjust the "Aggregate" and "Response" nodes if additional metadata fields are needed.
   • Modify the "Output" node to format the response (e.g., changing Sources: {{...}} to match your preferred style).

4. Testing

   • Trigger the workflow manually to test document ingestion.
   • Use the chat interface to verify responses include accurate source attribution.

Note: Replace placeholder values (e.g., QDRANTURL) with actual endpoints before deployment.

Need help customizing?

Contact me for consulting and support or add me on Linkedin.