Production-Grade Agentic RAG System: An Insight into Intelligent Retrieval-Augmented Generation Architecture via Paper Management

This open-source project demonstrates a complete implementation of a production-grade Agentic RAG system, using academic paper management as the scenario, covering the full-stack technology stack including data ingestion, parsing, indexing, retrieval, RAG workflow, intelligent agent workflow, and observability. By integrating the reasoning and planning capabilities of intelligent agents, Agentic RAG addresses the limitations of basic RAG when handling complex queries, enabling autonomous retrieval strategy decision-making, information quality assessment, and iterative answer optimization.

Retrieval-Augmented Generation (RAG) has become the mainstream architecture for large language model applications, but basic RAG only solves the 'knowledge update' problem and often struggles with complex queries. By introducing the reasoning and planning capabilities of intelligent agents, Agentic RAG elevates RAG to a new level—the system can not only retrieve information but also independently decide retrieval strategies, evaluate information quality, and iteratively optimize answers.

Data Ingestion Layer: Airflow-Orchestrated ETL Pipeline

The project uses Apache Airflow to implement automated data ingestion, including scheduled new paper source fetching, fault-tolerant retry mechanisms, incremental update processing, and task dependency management, ensuring pipeline reliability and maintainability.

Document Parsing: PDF to Structured Text

Implements a multi-level parsing strategy: PDF text extraction (including OCR), structure recognition (titles/abstracts/chapters, etc.), metadata extraction (authors/dates/keywords), and table/formula processing.

Indexing Layer: OpenSearch Hybrid Retrieval

Uses OpenSearch to support dense vector retrieval, sparse vector retrieval (BM25), and hybrid retrieval, combined with an intelligent document splitting strategy to balance context integrity and retrieval accuracy.

RAG Core: FastAPI-Powered Inference Service

Builds asynchronous APIs based on FastAPI, supporting multiple retrieval configurations (Top-K/similarity threshold/re-ranking), optimized prompt templates, intelligent context assembly, and SSE streaming responses.

Agentic Layer: Autonomous Decision-Making Retrieval Agent

A core innovation: the agent can perform query analysis, multi-step retrieval, information verification, iterative optimization, and tool calling to achieve autonomous resolution of complex problems.

Model Service: Ollama Local Inference

Integrates Ollama to support local open-source models (Llama/Mistral, etc.), providing model management, GPU acceleration, privacy protection, and cost optimization capabilities.

Observability: Langfuse Full-Stack Tracing

Integrates Langfuse to implement request tracing, performance monitoring, cost analysis, and quality assessment, ensuring observability of the production system.

Data Persistence: PostgreSQL Multi-Purpose Storage

PostgreSQL serves roles such as metadata storage, vector storage (pgvector extension), session management, and audit logging.

User Interaction: Telegram Bot Integration

Provides a Telegram Bot interface, supporting natural language queries, paper recommendations, abstract generation, and in-depth Q&A interactions.

Production environment considerations include:

• Scalability: Microservices architecture, stateless design, message queue asynchronous processing;
• Reliability: Multi-layer fault tolerance, data backup and recovery, canary release and rollback;
• Security: Input validation and filtering, access control, sensitive data encryption;
• Maintainability: Comprehensive log monitoring, externalized configuration, complete documentation.

For developers building production-grade RAG systems, the project provides:

1. Architecture Reference: Demonstrates how components are organically integrated;
2. Technology Selection: Explains the reasons for choosing specific technology stacks;
3. Best Practices: Contains a wealth of engineering practice details;
4. Extension Foundation: Can serve as a starting point for customized development.

Agentic RAG represents the next stage of RAG technology development. This open-source project, through the academic paper management scenario, demonstrates how to transform the Agentic RAG concept into a production-ready system, providing valuable references to the community in terms of technical architecture design, component selection, and engineering practices. As LLM applications deepen, end-to-end solutions will become increasingly important.