The author details Kimi Linear's architecture, training challenges, aggressive MoE sparsity, hybrid linear attention design, benchmark gains, and post‑training insights, offering a transparent technical review of this 48B‑parameter MoE LLM built on 5.7 T tokens.
This article introduces the Model Context Protocol (MCP) and Large Language Models (LLM), explains their core concepts, transmission mechanisms, lifecycle, and essential modules, and provides step‑by‑step code examples for creating an MCP server, adding tools, resources, prompts, and debugging workflows to accelerate AI‑driven development.
This article explains the evolution from RNNs through Encoder‑Decoder models to Transformers, detailing self‑attention, multi‑head attention, and masked attention, and then describes what Large Language Models are, their key components, capabilities, limitations, and common applications.
This article analyzes two contrasting AI agent design paradigms—process‑driven workflow orchestration and autonomous agent intelligence—examines their strengths and limitations, and proposes a hybrid architecture that fuses deterministic pipelines with dynamic planning, tool use, and memory mechanisms to achieve both reliability and adaptability.
This guide shows how to evolve a simple Retrieval‑Augmented Generation prototype into a production‑grade system by strengthening data ingestion, optimizing retrieval with hybrid and reranking techniques, adding query rewriting, long‑context handling, reinforcement learning, and multimodal support, so candidates can demonstrate real engineering depth in interviews.
This article summarizes six recent arXiv papers that explore how large language models, graph‑theoretic methods, generative frameworks, hypergraph multimodal architectures, GroupSHAP‑enhanced forecasting, and multi‑agent LLM workflows can improve financial signal extraction, portfolio optimization, and stock‑price prediction, providing empirical results on S&P 500 data.
RIVAL (Reinforcement Learning with Iterative and Adversarial Optimization) introduces an adversarial game between a reward model and a translation LLM, combining qualitative preference rewards with quantitative metrics like BLEU, to overcome distribution shift in RLHF and achieve superior performance on conversational subtitle and WMT translation tasks.
Meta’s recent paper reveals a sigmoid‑shaped scaling law for LLM reinforcement learning, presents extensive 40‑k GPU‑hour experiments, compares various RL designs such as PPO‑off‑policy‑k and Pipeline‑RL‑k, and distills the findings into a practical “ScaleRL” recipe that improves performance and efficiency.
This article shares the practical lessons learned from building Alibaba Cloud’s digital employee "YunXiaoEr Aivis", explaining why large‑language‑model agents often miss expectations and presenting ten concrete strategies—ranging from clear prompt design to memory management—that dramatically improve multi‑agent reliability.
LangChain’s stateless LLMs require a memory component to retain conversation context, and this article explains the importance of memory, compares built‑in memory types, and walks through two practical examples—custom buffer memory and basic message history—showing how to implement and use them with code.
LangGraph extends LangChain by letting developers define stateful, multi‑agent LLM workflows as graphs with nodes, edges, and shared state, and the article walks through core concepts, typical use cases, and a detailed example that shows how to define state, nodes, edges, compile and run the graph.
Learn how LangChain enables multimodal interactions by preparing inputs, constructing prompts, invoking models like GPT‑4o, and processing responses, with a complete example that demonstrates image‑question answering, code walkthrough, environment setup, and key considerations for API keys and image URLs.
This guide explains what LangChain tools are, why clear descriptions matter, and walks through three ways to create them—using the @tool decorator, StructuredTool with Pydantic models, and custom BaseTool subclasses—plus examples of built‑in tools and reference links.
This article explains the core concept of LangChain agents—combining an LLM, a set of tools, and a reasoning‑action loop—to enable dynamic decision‑making, tool invocation, and iterative observation for solving complex, multi‑step tasks.
The article examines the fundamental similarity between SFT and RL loss functions for large language models, explains why RL training is prone to instability, discusses infrastructure and data quality challenges, and reviews practical tricks and reward‑model considerations for more reliable RL fine‑tuning.
BiomedSQL and LogicCat are two newly released Text‑to‑SQL datasets that challenge large language models with complex biomedical reasoning, multi‑step logical inference, and domain‑specific knowledge, offering detailed analyses of query types, scientific reasoning categories, and performance gaps that highlight current LLM limitations.
This article explains how proper document chunking—choosing the right chunk size, overlap, and structure‑aware boundaries—directly impacts the relevance, factuality, and efficiency of Retrieval‑Augmented Generation pipelines, and provides multiple Python implementations ranging from simple fixed‑length splits to semantic and hybrid approaches.
This article analyzes the architecture, core capabilities, and workflow orchestration of LLM‑native application platforms Tasking AI and Dify, comparing their microservice designs, plugin management, multi‑tenant isolation, and GraphEngine execution to highlight strengths, trade‑offs, and future development trends.
Exploring Bilibili’s large‑scale data platform, this article details its five‑layer, storage‑compute separated architecture, the massive daily workload of offline and real‑time tasks, common failure and slowdown causes, and how an LLM‑powered intelligent assistant is being developed to help engineers troubleshoot efficiently.
Researchers from Texas A&M and UT Austin demonstrate that prolonged pre‑training on low‑quality, short‑form web content causes large language models to suffer irreversible cognitive decline—manifested as attention loss, broken reasoning chains, and personality distortion—highlighting data quality as a critical training‑time safety issue.
Over the past year, generative recommendation has made substantial progress by leveraging large language models' powerful sequence modeling and reasoning abilities, introducing a new paradigm that replaces complex handcrafted features, addresses traditional recommendation bottlenecks, and outlines the evolution, core technologies, engineering challenges, and future directions of LLM‑based recommendation systems.
The article explains how traditional prompt engineering is giving way to Context Engineering and the Agentic Context Engineering (ACE) framework, which lets large language model agents continuously learn and improve through evolving, well‑structured context without fine‑tuning.
Bilibili’s game algorithm team built a four‑layer, LLM‑based translation platform that automates terminology extraction, retrieval‑augmented generation, and quality assessment, dramatically reducing localization cycles by over 85% and costs by up to 80% while supporting ten languages and ensuring consistent, culturally‑accurate game text.
This article explains how to design and optimize the generation module of Retrieval‑Augmented Generation systems by building robust prompts, merging multi‑source information, controlling answer formats, and applying post‑generation verification to reduce hallucinations and improve enterprise‑grade performance.
Zero-shot prompting lets large language models perform tasks without examples, and by following principles of clarity and structured instructions, advanced strategies such as emotion prompting, zero-shot chain-of-thought, RE2 re-reading, Rephrase-and-Respond, role-play, and System-2 Attention can significantly improve accuracy and response quality across translation, reasoning, and QA tasks.
PacketScope leverages eBPF to trace every packet inside the TCP/IP stack, visualizes protocol interactions, and uses large language models to automatically detect and block sophisticated network attacks with zero‑delay, addressing the growing $10.5 trillion cyber‑crime threat projected for 2025.
This article explains the Model Context Protocol (MCP) – an open, universal standard that lets large language models seamlessly interact with external tools and data – covering its core architecture, why it’s needed, underlying principles, tool‑selection mechanics, a step‑by‑step Python server implementation, and practical usage tips.
This article breaks down the end‑to‑end engineering pipeline for constructing a Retrieval‑Augmented Generation (RAG) knowledge base, covering document parsing, data cleaning, semantic chunking, embedding, and index creation, plus practical optimization tips and a concise interview answer framework.
This digest presents seven recent arXiv papers that explore large‑language‑model‑driven portfolio scoring, hybrid ResNet‑RMT covariance denoising for crypto, LLM‑enhanced financial causal analysis, multilingual news alignment for stock returns, three‑step bubble prediction with news and macro data, multimodal volatility forecasting, and news‑aware reinforcement trading, each with reported performance gains.
The article outlines three post‑RAG knowledge‑engineering approaches—In‑Context Learning with dynamic few‑shot selection, Online Learning encompassing Meta‑Learning and Lifelong Learning to quickly adapt to new tasks, and the Small Language Model path that combines fine‑tuned task‑specific experts with LLM‑SLM collaboration for efficient, privacy‑preserving inference.
This article explains the fundamental differences between workflows and agents, introduces seven core design patterns—including three workflow patterns and four agent patterns—provides Python examples using Ollama, and shows how to combine these patterns to create robust, autonomous AI applications.
This article explains why most LLM‑based agents only perform pseudo‑planning through prompts or hard‑coded loops, outlines when to rely on prompt‑driven versus program‑driven planning, compares popular frameworks such as ReAct, MRKL, BabyAGI and AutoGPT, and clarifies what true autonomous planning would require.
This guide walks through the complete end‑to‑end process of training a large language model from scratch, covering data collection, cleaning, tokenization, pre‑training objectives and engineering, post‑training alignment methods, scaling laws, over‑fitting mitigation, and gradient‑stability techniques.
This tutorial revisits a simple text‑generation function, explains how temperature scaling and top‑K sampling reshape token probability distributions, demonstrates their effects with PyTorch code and visualizations, and shows how to integrate both techniques into an improved generation routine for more diverse and human‑like outputs.
An LLM-driven intelligent data-ingestion framework replaces manual, months-long integration with an automated code-generation and execution loop that auto-recognizes schemas, maps structures, extracts quality rules, builds deployment packages, cutting onboarding time from three months to three days while eliminating human effort.
This article walks through the design and implementation of a two‑layer memory architecture for LangGraph agents, covering short‑term and long‑term stores, various storage back‑ends, prompt engineering, utility functions, node definitions, human‑in‑the‑loop interrupt handling, and how user feedback is captured and used to continuously update the agent’s behavior.
This article revisits nanochat's core components, detailing the preparation of diverse training datasets, the scaling calculations for tokens and parameters, the model's MQA and KV‑cache design, the full training pipeline with gradient accumulation and mixed‑precision, cost breakdown, inference optimizations, evaluation tasks, and identified limitations with suggested improvements.
This article introduces Google’s Tunix library for JAX‑based LLM post‑training, explains its core features such as supervised fine‑tuning, reinforcement learning and knowledge distillation, and provides detailed installation steps and a complete TPU‑accelerated QLoRA fine‑tuning workflow on the Gemma 2B model, including code snippets and inference testing.
This article analyzes the rStar2‑Agent paper, revealing how Agentic Reinforcement Learning, the GRPO‑RoC algorithm, a high‑throughput code‑execution service, and a three‑stage training recipe let a modest 14‑billion‑parameter model surpass much larger LLMs on challenging math benchmarks.
This guide walks through deploying Alibaba's open‑source Tongyi‑DeepResearch 30B MoE model locally, configuring FastAPI and A2A interfaces, implementing a ReAct‑style agent with LangGraph, setting up research tools, and testing the full UI‑API‑Agent pipeline via CLI and Streamlit.
This article details the implementation of an LLM‑based code‑review agent in a C3‑level secure repository, describing its RAG‑enhanced knowledge base, CI pipeline integration, real‑world results, prompt engineering, and ongoing optimization to boost review efficiency and defect detection.
QuantAgent introduces a multi‑agent LLM framework that replaces text‑based inputs with raw OHLC price signals, decomposes trading decisions into Indicator, Pattern, Trend, Risk, and Decision agents, and achieves substantially higher direction accuracy and returns across ten financial assets in zero‑shot HFT experiments.
This article provides a thorough analysis of nanochat’s source code, detailing transformer component differences, precise parameter‑size formulas, FlashNorm and ReLU² innovations, scaling‑law insights, memory‑usage estimations, and the distributed optimizer and training pipelines used to build the model.
A recent study reveals that increasingly impolite prompts can significantly improve large language model accuracy, challenging common assumptions about politeness and prompting while offering practical insights for effective AI interaction.
This article explores how Spring AI Alibaba enables the development of autonomous AI agents that run on schedules, interact with humans when needed, and handle tasks such as periodic business automation, batch processing, emergency response, and long‑cycle data analysis, illustrated with Java code examples.
This tutorial shows how to let large language models autonomously invoke external services—specifically a MySQL database—by using the Model Context Protocol (MCP), covering environment setup, MCP service implementation, agent integration, and real‑world execution results.
This presentation by iQIYI’s Technical Director Zhang Xiaoming details the evolution of BI systems, introduces the Chat BI framework, explains its three‑step implementation, outlines architectural design, data‑warehouse integration, performance optimizations, and user‑operation strategies, revealing how AI and RAG empower smarter data analytics.
This article walks through the DeepSeek‑V3.2‑Exp inference codebase, detailing its MoE architecture, Multi‑Head Latent Attention, FP8 quantization, custom CUDA kernels, and 8‑GPU NCCL‑based distributed execution from initialization through prefill and decode stages.
This article presents an LLM‑driven intelligent data platform ingestion solution that automates schema recognition, mapping, quality rule extraction, and package building, reducing integration cycles from three months to three days while eliminating manual effort and enhancing scalability and control.
The Model Context Protocol (MCP) introduced by Anthropic addresses the scalability, integration, and context‑transfer limitations of traditional Function Calling by offering a standardized, bidirectional, and context‑aware communication layer that simplifies tool discovery, security, and workflow orchestration for LLM‑driven agents.
Large language models are evolving from free‑form text generators to reliable data providers by mastering structured output through prompt engineering, validation frameworks, constrained decoding, supervised fine‑tuning, reinforcement learning, and API‑level capabilities, enabling seamless integration with software systems while addressing hallucinations and format reliability.
TS‑Agent is a modular LLM‑driven framework that formalizes financial time‑series modeling as a three‑stage iterative decision process, leveraging structured knowledge bases, dynamic memory, and a feedback‑driven code‑editing loop to outperform AutoML baselines in accuracy, robustness, and auditability.
CollabLLM tackles the limitations of large language models in everyday multi‑turn dialogues by introducing a user‑centric, multi‑turn training framework that leverages simulated interactions, multi‑round reward modeling, and veRL toolchain support, achieving superior performance over single‑turn baselines.
Ant Group's open‑source dInfer framework dramatically speeds up diffusion language model inference—achieving more than a ten‑fold boost over Fast‑dLLM, surpassing autoregressive baselines, and delivering 1011 tokens per second on HumanEval—by tackling computational cost, KV‑cache invalidation, and parallel decoding challenges through modular system‑level innovations.
This article explains how to design a high‑performing AI Agent by constructing a robust knowledge base, orchestrating efficient workflows, and crafting precise prompts, covering vector storage, graph databases, retrieval strategies, and practical prompt‑engineering techniques.
This article introduces the BIRD-CRITIC benchmark, a comprehensive SQL diagnostic dataset spanning multiple dialects, evaluates large language models' ability to repair real-world database queries, and discusses its design, multi‑dialect support, data quality processes, and experimental results.
This article outlines practical techniques—including prompt engineering, domain fine‑tuning, retrieval‑augmented generation, structured outputs, workflow constraints, model parameter control, behavior rules, risk‑based AI participation, and comprehensive governance—to curb the unpredictability of large language model agents in enterprise settings.
This article reviews Trading‑R1, an open‑source LLM inference framework that integrates multimodal financial data, three‑stage supervised‑fine‑tuning and reinforcement learning to generate structured investment arguments and risk‑adjusted trade decisions, achieving superior Sharpe ratio and drawdown performance on real‑world stock and ETF tests.
This article reviews Kuaishou's two‑year exploration of multimodal large‑model techniques for advertising, outlining challenges in content‑domain ad estimation, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together improve ad system performance.
This article explores the essential stop‑condition designs for AI agents, detailing hard limits, task‑completion checks, explicit termination tools, loop detection, error accumulation, and user interruption, and then examines concrete implementations in OpenManus and Gemini CLI with code examples and multi‑layer safeguards.
This guide walks you through upgrading Dify using Docker Compose or source code deployment, running required migration commands, backing up data, and fixing the "Invalid context structure" error caused by iterator output changes in version 1.9.1, with detailed code snippets and troubleshooting steps.
Learn how to use LangChain's Document Loaders to import data from files, web pages, databases, and APIs, understand the Document object structure, compare load() versus lazy_load(), and follow a step‑by‑step Python example that demonstrates loading, inspecting, and optionally processing documents with an LLM.
This article explores Tencent’s large language model deployments across diverse business scenarios—content generation, intelligent customer service, and role‑playing—detailing the underlying RAG, GraphRAG, and Agent technologies, their principles, practical implementations, and the advantages they bring to enterprise AI solutions.
Spring AI Alibaba (SAA) provides a robust framework for building autonomous, scheduled AI agents that can operate independently, respond to events, and involve human oversight, enabling use cases such as automated business reporting, batch data processing, emergency response, and sentiment analysis, with detailed code examples and deployment guidance.
This article surveys the most influential large language models released since 2017, detailing the core innovations of Transformer, BERT, GPT series, T5, Retrieval‑Augmented Generation, and the latest LLaMA and Meta models, while highlighting their architectures, training paradigms, and impact on NLP research.
This guide explains why custom output parsers are needed for LangChain when dealing with non‑JSON or XML responses, walks through inheriting BaseOutputParser, implementing parse() and optional format instructions, and provides a complete Python example that converts a simple "Key: Value" string into a dictionary.
This tutorial demonstrates how to integrate multiple large language models from different providers into a Spring Boot application using Spring AI, configure primary, secondary, and tertiary models, and implement a fallback mechanism with Spring Retry to ensure high availability of the chatbot.
This article examines the Language Self‑Play (LSP) approach for data‑free training of large language models, detailing its challenger‑solver game formulation, advantage calculations, loss functions, self‑reward extension, experimental setup on AlpacaEval, and results that show LSP can match or surpass data‑driven baselines.
This article explains Retrieval‑Augmented Generation (RAG), an AI framework that combines external knowledge retrieval with large language models, covering its motivations, data preparation, chunking strategies, vectorization, storage, query processing, retrieval, reranking, prompt engineering, and LLM generation, plus practical optimization tips.
Ling-1T, a trillion‑parameter flagship non‑thinking model, combines 50 billion active parameters per token, 128 K context, Evo‑CoT reasoning, and FP8 mixed‑precision training to achieve state‑of‑the‑art performance on complex reasoning, code generation, and multimodal tasks while outlining its architecture, benchmarks, limitations, and future roadmap.
This article explains the definitions, technical implementations, functions, limitations, and collaborative workflow of short‑term and long‑term memory in large‑language‑model agents, detailing context windows, attention mechanisms, vector storage, retrieval strategies, and future research directions for building personalized, continuously learning AI agents.
The article analyzes the security challenges introduced by the open Model Context Protocol (MCP) ecosystem, outlines typical attack vectors such as command‑execution hijacking and indirect prompt injection, and presents MCPScan—an open‑source scanner that combines static taint analysis with LLM‑driven reasoning to detect exploitable tool chains before deployment.
The paper introduces TradingGroup, a five‑agent LLM‑based quantitative trading framework that incorporates a self‑reflection mechanism, dynamic risk management, and an automated data‑synthesis pipeline, and demonstrates superior cumulative returns, Sharpe ratios, and lower drawdowns than rule‑based, ML, RL, and existing LLM strategies on five real‑world stock datasets.
This article explains how an AI‑driven multi‑agent platform automates fault postmortem generation, enriches analysis with memory management, prompt engineering, and RAG techniques, and delivers actionable insights for SREs, developers, and non‑technical stakeholders, ultimately shifting incident handling from reactive to proactive.
This article explains how the browser‑use/qa‑use platform leverages large language models to let testers describe test cases in natural language, automatically generates browser actions, executes them, and provides detailed reports, dramatically reducing script maintenance and boosting testing efficiency.
This article reviews Kuaishou’s two‑year exploration of large‑model techniques in advertising, detailing the content‑domain estimation challenges, how multimodal and LLM approaches improve full‑domain behavior utilization and external knowledge integration, and introducing the COPE product‑content representation framework and the LEARN LLM knowledge‑transfer system.
This article explains why LLMs output plain text, introduces LangChain output parsers as the bridge to structured data, details their workflow, reviews built‑in parsers, and walks through a complete Python example that builds a prompt‑model‑parser chain to generate a JSON‑based joke.
The article explains the differences between AI Agents and AI Workflows, compares their characteristics, introduces the hybrid Agentic Workflow concept, and offers practical recommendations for building enhanced LLM applications using simple prompts or advanced frameworks.
Bilibili’s massive video platform relies on a five‑layer, storage‑compute separated big‑data architecture, handling hundreds of thousands of daily tasks, and now leverages large‑language‑model assistants to automatically diagnose and resolve frequent task failures and performance slowdowns.
This article explains why prompt templates are essential for LLM interactions, compares LangChain's PromptTemplate and ChatPromptTemplate, and walks through a complete Python example that prepares the environment, builds reusable templates, formats messages, and integrates them into an LCEL chain for chat models.
This article explains why and how to create custom LLM and chat model wrappers in LangChain by inheriting base classes, implementing core methods like _call and _generate, and demonstrates a full example with an echo‑based chat model, including streaming support and testing benefits.
This article explains how function (tool) calling lets large language models like GPT or Gemini invoke external APIs, walks through defining tools with LangChain, and demonstrates a complete Python example that fetches real‑time weather data and returns a natural‑language answer.
This article analyzes a recent paper that unifies supervised fine‑tuning (SFT) and reinforcement learning (RL) for large language models under a single gradient estimator, introduces the Unified Policy Gradient Estimator (UPGE) and the Hybrid Post‑Training (HPT) algorithm, and demonstrates their superior performance on math reasoning benchmarks.
This article explains LangChain's two core abstractions—LLM for simple text completion and ChatModel for multi‑turn conversational AI—detailing their input/output formats, practical code examples, and why ChatModel is generally preferred for modern dialogue applications.
This article outlines Kuaishou's two‑year exploration of large‑model techniques in advertising, detailing challenges of sparse cross‑domain data, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together improve ad system effectiveness.
A recent interdisciplinary study from Caltech, Cambridge and others shows that while large language models can present idealized personalities on questionnaires, their actual behavior in tasks diverges sharply, exposing a ‘personality illusion’ that challenges current AI alignment approaches.
Over the past two years, Kuaishou has leveraged multimodal large‑model techniques to overcome sparse advertising data, integrating full‑domain user behavior and external knowledge via the COPE unified product representation framework and the LEARN LLM knowledge‑transfer system, achieving measurable business gains.
Learn how to serialize LangChain components—including prompts, chains, and agents—using JSON and YAML, enabling reproducibility, collaboration, persistence, and decoupling, with step‑by‑step code examples for dumping objects to files and loading them back into executable LLM pipelines.
AlphaAgents introduces a role‑based multi‑agent framework—Fundamental, Sentiment, and Valuation agents—leveraging LLMs to analyze 10‑K reports, news, and price data, with a debate mechanism via Microsoft AutoGen; experiments on 15 tech stocks show superior cumulative returns and Sharpe ratios under risk‑neutral and risk‑averse settings compared to single‑agent baselines.
This article explains the nature, types, and causes of hallucinations in large language models used for NL2SQL, reviews both unsupervised and supervised detection methods, and introduces an efficient token‑confidence based Active Sampling Detection (ASD) approach with practical deployment examples and future research directions.
This article details the engineering practice of constructing a complete data infrastructure for large‑language‑model (LLM) applications using Alibaba Cloud SLS, covering the observability challenges of the Dify platform, the redesign of the architecture, and the resulting improvements in monitoring, diagnosis, and quality optimization.
LangChain’s callback system lets developers hook into every stage of an LLM chain— from chain start/end to token generation—using built‑in handlers like StdOutCallbackHandler or custom handlers derived from BaseCallbackHandler, with examples showing constructor‑level and request‑level attachment, plus a custom handler implementation.
This article examines the challenges of processing massive multimodal data in enterprises and presents a knowledge‑augmentation framework that leverages Retrieval‑Augmented Generation, memory‑inspired architecture, and feedback loops to enable reliable, scalable AI‑driven decision making across diverse business scenarios.
This article explains Bilibili's implementation of a large‑language‑model based intelligent assistant, detailing the platform's five‑layer architecture, the huge volume of offline and real‑time jobs, common user issues like task failures and slowdowns, and how AI can help automate troubleshooting.
This article explains why initial retrieval in Retrieval‑Augmented Generation often yields noisy results, describes how rerankers act as quality filters to improve relevance, compares the leading 2025 reranker models—including Cohere, bge‑reranker, Voyage, Jina, FlashRank, and MixedBread—and provides code snippets, evaluation metrics, and guidance for selecting the right model for specific use cases.
A lightweight, real‑time platform called Awesome AI Agents Live, built by East China Normal University students, aggregates the latest AI Agent research, offering targeted navigation, structured classification, multi‑dimensional sorting, and concise paper summaries for researchers, engineers, and students.
JD Cloud has released JoyAgent‑JDGenie, the first fully product‑grade open‑source Java multi‑agent system that bundles front‑end, back‑end, framework, engine and core agents, supports major LLMs, offers layered architecture, Docker or manual deployment, and showcases demos such as PPT generation and sales analysis.
This article explains Retrieval‑Augmented Generation (RAG), an AI framework that combines traditional information retrieval with large language models, detailing its core workflow—from knowledge preparation, chunking, and embedding to vector database storage and the question‑answering stage—while highlighting key challenges, tools, and optimization strategies.
This article explains Retrieval‑Augmented Generation (RAG), an AI framework that combines traditional information retrieval with large language models, covering its core workflow—from knowledge preparation, data cleaning, and metadata extraction to query preprocessing, vector retrieval, reranking, information integration, and final LLM generation, while also reviewing common embedding models and vector databases.
This article breaks down the GPU memory requirements of large language models during training and inference, detailing the contributions of model weights, optimizer states, activations, KV cache, and activation recomputation, and provides concrete formulas, examples, and scaling insights for models like Qwen3 and DeepSeek V3.
