Cezzis Cocktail RAG System: End-to-End Intelligent Retrieval-Augmented Generation Workflow

This article introduces cezzis-com-ingestion-agentic-wf—a domain-specific end-to-end RAG system for cocktails, providing intelligent search and Q&A capabilities for cezzis.com. Built on Python, Qdrant vector database, and Ollama local large model, combined with Azure Cosmos DB data source and E5 embedding model, it offers REST API services for semantic search and conversational Q&A. Its core value lies in improving answer accuracy, timeliness, and traceability through retrieval-augmented generation, while integrating agent capabilities to optimize user experience.

RAG (Retrieval-Augmented Generation) is an AI architecture combining information retrieval and text generation, divided into three stages: retrieval, augmentation, and generation. Compared to pure generative models, RAG has significant advantages: reducing hallucinations based on real data, ensuring timeliness via dynamically updatable knowledge bases, enabling traceability of answers to sources, and lowering costs by eliminating the need for fine-tuning large models. This system addresses the needs of cocktail enthusiasts, solving the problem that traditional search struggles to handle natural language queries and professional knowledge Q&A.

Core Technology Stack: Python backend (asynchronous frameworks like FastAPI/Flask), Qdrant vector database (vector storage, semantic search), Ollama local large model (text generation, embedding, privacy protection), Azure Cosmos DB (cocktail data storage), E5 embedding model deployed via TEI service (excellent semantic similarity performance).

System Architecture is divided into two phases: Data Preparation (extract data from Cosmos DB → document chunking → embedding generation → vector storage in Qdrant); Online Service (query embedding → semantic retrieval → context construction → answer generation via Ollama).

This system integrates agent capabilities, differing from traditional single-step RAG processes: Query Understanding (judging multi-step retrieval needs), Tool Calling (using tools like search/computation on demand), Iterative Optimization (adjusting retrieval strategies), Self-Correction (correcting errors). In the cocktail scenario, this manifests as: multi-round retrieval (e.g., first searching for summer cocktails then low-alcohol ones), reasoning ability (recommending recipes based on available ingredients), and clarification interaction (proactively asking for preferences when queries are ambiguous).

Core API Endpoints: Semantic Search (POST /api/search, supports filtering and top_k), Conversational Q&A (POST /api/chat, supports streaming output), Ingredient Recommendation (POST /api/recommend, based on available ingredients).

Deployment uses Docker Compose orchestration (API service, Qdrant, TEI, Ollama), supports incremental data synchronization (scheduled updates from Cosmos DB), and includes monitoring (request latency, retrieval quality, generation quality) and log alerts.

Application Scenarios: Website search enhancement (natural language queries like "pink cocktails suitable for women"), virtual bartender assistant (recipe guidance, ingredient substitution, cultural background), content creation assistance (auto-generate descriptions, recommend topics).

Technical Highlights: Modular design (separated components for easy expansion), local-first approach (Ollama/TEI local inference for privacy protection), production-ready (complete error handling and monitoring), extensible to other domains (just replace the data source).

cezzis-com-ingestion-agentic-wf is an excellent practical case of RAG systems, integrating modern AI technologies to provide practical knowledge services. It offers a clear architectural reference for developers, demonstrating how to build domain-specific intelligent Q&A systems. As RAG technology matures, such systems will be widely applied in more industries like food and tourism, providing users with more accurate and personalized knowledge services.