My solution for the "Agentic Arena Community Contest" (RAG, Qdrant, Mistral OCR)

🤖📈 This workflow is my personal solution for the Agentic Arena Community Contest, where the goal is to build a Retrieval-Augmented Generation (RAG) AI agent capable of answering questions based on a provided PDF knowledge base. --- Key Advantages ✅ End-to-End RAG Implementation Fully automates the ingestion, processing, and retrieval of knowledge from PDFs into a vector database. ✅ Accuracy through Multi-Layered Retrieval Combines embeddings, Qdrant search, and Cohere reranking to ensure the agent retrieves the most relevant policy information. ✅ Robust Evaluation System Includes an automated correctness evaluation pipeline powered by GPT-4.1 as a judge, ensuring transparent scoring and continuous improvement. ✅ Citation-Driven Compliance The AI agent is instructed to provide citations for every answer, making it suitable for high-stakes use cases like policy compliance. ✅ Scalability and Modularity Can easily integrate with different data sources (Google Drive, APIs, other storage systems) and be extended to new use cases. ✅ Seamless Collaboration with Google Sheets Both the evaluation set and the results are integrated with Google Sheets, enabling easy monitoring, iteration, and reporting. ✅ Cloud and Self-Hosted Flexibility Works with self-hosted Qdrant on Hetzner, Mistral Cloud for OCR, and OpenAI/Cohere APIs, combining local control with powerful cloud AI services. --- How it Works 1. Knowledge Base Ingestion (The "Setup" Execution): When started manually, the workflow first clears an existing Qdrant vector database collection. It then searches a specified Google Drive folder for PDF files. For each PDF found, it performs the following steps: Uploads the file to the Mistral AI API. Processes the PDF using Mistral's OCR service to extract text and convert it into a structured markdown format. Splits the text into manageable chunks. Generates embeddings for each text chunk using OpenAI's model. Stores the embeddings in the Qdrant vector store, creating a searchable knowledge base. 2. Agent Evaluation (The "Testing" Execution): The workflow is triggered by an evaluation Google Sheet containing questions and correct answers. For each question, the core AI Agent is activated. This agent: Uses the RAG tool to search the pre-populated Qdrant vector store for relevant information from the PDFs. Employs a Cohere reranker to refine the search results for the highest quality context. Leverages a GPT-4.1 model to generate an answer based strictly on the retrieved context. The agent's answer is then passed to an "LLM as a Judge" (another GPT-4.1 instance), which compares it to the ground truth answer from the evaluation sheet. The judge provides a detailed score (1-5) based on factual correctness and citation accuracy. Finally, both the agent's answer and the correctness score are saved back to a Google Sheet for review. --- Set up Steps To implement this solution, you need to configure the following components and credentials: 1. Configure Core AI Services: OpenAI API Credentials: Required for the main AI agent, the judge LLM, and generating embeddings. Mistral AI API Credentials: Necessary for the OCR service that processes PDF files. Cohere API Credentials: Used for the reranker node that improves retrieval quality. Google Service Accounts: Set up OAuth for Google Sheets (to read questions and save results) and Google Drive (to access the PDF source files). 2. Set up the Vector Database (Qdrant): This workflow uses a self-hosted Qdrant instance. You must deploy and configure your own Qdrant server. Update the Qdrant Vector Store and RAG nodes with the correct API endpoint URL and credentials for your Qdrant instance. Ensure the collection name (agentic-arena) is created or matches your setup. 3. Connect Data Sources: PDF Source: In the "Search PDFs" node, update the folderId parameter to point to your own Google Drive folder containing the contest PDFs. Evaluation Sheet: In the "Eval Set" node, update the documentId to point to y