Building a Professional RAG System for Publications

Introduction

Retrieval Augmented Generation (RAG) has revolutionized how we interact with large document collections. This publication details the evolution of the RAG Publications Assistant—from a static document retriever to a professional, session-isolated, and persistent AI platform.

The Challenge: Context Leakage in RAG

A common failure in multi-document RAG systems is Context Leakage. If a user switches from "Research Paper A" to "Research Paper B," the AI often retains "memory" of the previous document, leading to hallucinations or mixed-context answers.

Our system solves this through a dual-isolation strategy:

1. Vector Isolation: Every upload triggers the creation of a unique, transient Qdrant collection.
2. Memory Isolation: We use uuid.uuid4() to generate fresh chat_id markers, ensuring the AI only retrieves past questions relevant to the current active document.

Core Architectural Pillars

1. Nomic Embedding Optimization (Task-Specific)

We migrated from Jina to Nomic AI (nomic-embed-text-v1) to leverage task-specific embedding types, which significantly improve retrieval precision:

• Index Time (search_document): Optimizes the high-dimensional representation of the publication text for storage.
• Query Time (search_query): Tailors the user's question vector to match the document storage format, reducing noise in semantic search.

# Implementation of Task-Specific Retrieval
query_embedding = generate_nomic_embeddings_batch(
    API_KEY, 
    [refined_question], 
    task_type="search_query"
)[0]

2. Persistent Relational Memory (MySQL)

Unlike ephemeral session memory, we implemented a MySQL-backed persistent storage layer.

• Contextual Refinement: Before querying the vector DB, the system fetches the last 3 interactions to reformulate the user's question. This allows the AI to understand follow-up questions like "Why did the authors choose that method?"
• History Modal: Users can instantly access their long-term chat history via a professional modal, even after clearing their current active session.

3. Professional UX Design System

We moved beyond a simple chat box to a high-fidelity Glassmorphism-inspired UI:

• Dynamic Feedback Loop: Assistant responses now include thumbs-up/down icons, allowing for future RLHF (Reinforcement Learning from Human Feedback) integration.
• State Visibility: Real-time pulsing typing indicators and placeholder animations handle the perceived latency of LLM generations.
• Universal Theme Sync: A complete Light/Dark mode toggle implemented via CSS variables ensures a consistent visual experience across all system components.

Technical Specifications

Conclusion

The RAG Publications Assistant demonstrates that an effective academic tool requires more than just an LLM; it requires a robust infrastructure for memory persistence and context isolation. By grounding every answer in specific, user-controlled datasets and maintaining strict session boundaries, we provide a reliable platform for deep academic inquiry.

References

• Lewis, P., et al. (2020). "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks"
• Gao, L., et al. (2023). "Precise Zero-Shot Dense Retrieval without Relevance Labels"
• Nomic AI Documentation (2024). "Contrastive Learning for Task-Specific Embeddings"