Louis | strictfrog.com

Anxiety and Redemption in the Age of AI

Purely Fictional

Anxiety As an IT programmer, since last year, it seems I’ve fallen into an atmosphere of fearing unemployment in the AI era. My IT friends, the real IT environment, and social media flooded with A...

发布者 LuochuanAD，日期 2026-06-18

Enterprise AI Platform Architecture

Architecture

Background Since I live in Tokyo, I have encountered many companies that don’t know how to implement an AI platform. This article shares my experience. Design Approach (Brief Overview) ...

发布者 LuochuanAD，日期 2026-06-18

Implementing Frontend and Backend Service Separation with Nextjs and FastAPI

Building Efficient Web Interfaces

Background Previously, the Web pages for the AIAgent were built using Html+Css+Js. Besides looking unattractive, the session management and multi-user state handling in the Agent were overly co...

发布者 LuochuanAD，日期 2026-06-18

Design of the Memory Layer in ChatAgent

Memory Layer

Background In several previous ChatAgent projects, I encountered many challenges when designing the Memory Layer. In this article, I will explain how to design the Memory Layer effectively. C...

发布者 LuochuanAD，日期 2026-04-25

Evaluation Module of the RAG System and Memory Layer

Evaluation Module

Background In several previous AI Agent projects, I built RAG systems and Memory Layer. Now, I need to evaluate this RAG system and Memory Layer from the following dimensions: accuracy_score, p...

发布者 LuochuanAD，日期 2026-04-19

Resource Scheduling Strategies in AI Infrastructure Layer?

AI Infra Layer

Background In a privatized large model system, I have deployed Embedding models, Reranker models, and LLMs locally, and implemented high-concurrency batch processing. Configuration (config.py...

发布者 LuochuanAD，日期 2026-04-12

How to Build a High-Concurrency Module for Local Models?

AI Infra Layer

Background In a privatized large model system, I have deployed the Embedding model, Reranker model, and LLM locally, and I also need to support concurrency. Here, I have built a dedicated high-...

发布者 LuochuanAD，日期 2026-04-11

How to Deploy a Local Embedding Model?

AI Infra Layer

Background In a privatized large model system, to avoid external dependencies, reduce the risk of data leakage, and enhance controllability, the Embedding model is deployed locally. Additionall...

发布者 LuochuanAD，日期 2026-04-10

FastAPI+Uvicorn+Gradio

Rapidly Building High-Quality Web Interfaces

Background Previously, the web pages in AIAgent projects were built using Html+Css+Js, in one word: ugly. Choice FastAPI + Gradio for quickly building attractive UIs Below is the UI for...

发布者 LuochuanAD，日期 2026-04-04

How to build an AIOS architecture ?

architecture

Background Scenario: A lightweight ChatAgent embedded in a ChatGPT-style conversational interface, where users can engage in dialogue, document analysis, as well as other document generation an...

发布者 LuochuanAD，日期 2026-03-28

Prompt Template (Attack Prevention)

Prompt template

Background Prompt security and structure—this section provides only templates. Complete Prompt [System Prompt] [Tool Instructions] [Security Rules] [RAG Context] User Question System ...

发布者 LuochuanAD，日期 2026-03-22

Context Management and Filtering in Prompts

Technical Blog Title

Background When developing a ChatAgent for document analysis, in multi-turn conversations, users upload a file for analysis, and the LLM answers user questions based on the analysis results of ...

发布者 LuochuanAD，日期 2026-03-20

How to Survive in the Age of AI

Self-Reflection

How to Survive in the AI Era? In the AI era, job displacement is inevitable. This means your income source will no longer be stable, and your monthly disposable income will fluctuate. In this tr...

发布者 LuochuanAD，日期 2026-03-17

Data Quality Management in Fine-tuning

Data Cleaning

Background Fine-tuning model performance = 70% data structure design + 20% data quality + 10% training parameters This article focuses solely on data quality, specifically data cleaning. Com...

发布者 LuochuanAD，日期 2026-03-15

High-Quality SFT Data Structure Design for Fine-tuning

High-Quality SFT Data Structure

Background Fine-tuning model capability = 70% Data Structure Design + 20% Data Quality + 10% Training Parameters This article focuses solely on “Data Structure Design.” 5 Levels of High-Qual...

发布者 LuochuanAD，日期 2026-03-15

SFT Data Sources for Fine-tuning

Data Source Analysis

Background What are the sources of SFT data for fine-tuning LMs? 1 Open Source Datasets Common ones: Stanford University Alpaca https://raw.githubusercontent.com/tatsu-lab/stanford_alpaca/...

发布者 LuochuanAD，日期 2026-03-15

Simple Data Preparation and Preprocessing for Fine-tuning

Fine-tuning

Background For private LLMs, the most important part is fine-tuning the LLM with private data. This article first explains how to prepare and preprocess the data. Data Preparation By downloa...

发布者 LuochuanAD，日期 2026-03-15

Extreme Optimization of Latency in Private RAG

Private RAG

Background During the construction of a RAG system, I significantly improved the accuracy of knowledge retrieved by RAG through designing Query Rewrite, Rerank, Structural Chunking, Structural ...

发布者 LuochuanAD，日期 2026-03-14

Incremental Vector Update Strategy for Embedding

Avoid Duplicate Embeddings

Background Only embed “new or changed content” instead of reprocessing all documents each time. 1. Three Levels of Incremental Updates Level Granularity Recommenda...

发布者 LuochuanAD，日期 2026-03-14

The Design Philosophy of Self-Evolving

Self-Evolving

Background A self-improving (Self-R&D) Agent system that can continuously develop new tools, algorithms, strategies, and systems. Design Concept of a Self-Improving Agent Overall Archite...

发布者 LuochuanAD，日期 2026-03-14

The Design Philosophy of Self-Improving

Self-Improving

Background This article explains the design concept of the Self-Improving Agent, which adds self-improvement capabilities on top of the Autonomous Agent architecture. “Autonomous Agent archit...

发布者 LuochuanAD，日期 2026-03-13

Future AgentOS Core Architecture

AgentOS

Background Building the core architecture for future AgentOS based on the common failure reasons of most real-world agents Core Architecture of AgentOS State Machine Agent │ ...

发布者 LuochuanAD，日期 2026-03-12

Enterprise-level AIOS Architecture

AIOS

Background The goal of this architecture is to build a complete AI Operating System (AI OS) capable of: Integrating LLMs Managing Tools Managing Skills Managing Memory Supporting Agent...

发布者 LuochuanAD，日期 2026-03-11

AIAgent System Architecture of ChatGPT

Architecture

Background The previous article discussed the limitations of ChatGPT. This article aims to address the tasks ChatGPT cannot accomplish. Complete AI Agent System Architecture ...

发布者 LuochuanAD，日期 2026-03-10

ChatGPT Capability Limits

Optimizing Distributed Systems with ChatGPT for Techniques and Best Practices

Background Know yourself and your opponent to better develop private AI Agents 1. Innate Capabilities of ChatGPT (LLM Native Abilities) 1 Language Understanding & Generati...

发布者 LuochuanAD，日期 2026-03-08

AutoGPT Analysis and Autonomous Thinking

Auto-GPT

Background Auto-GPT is capable of formulating plans, parsing plans, breaking down plans, executing plans, evaluating execution results, and continuously iterating to produce outcomes while inte...

发布者 LuochuanAD，日期 2026-03-07

Design Philosophy of Workflow+limitedPlanning

Workflow + Limited Planning

Background In practical AI Agent systems, many teams adopt a hybrid architecture: Workflow + Limited Planning, where an LLM performs lightweight planning followed by executing a predefined work...

发布者 LuochuanAD，日期 2026-03-07

Dreams and Self-Reflection

Purely Fictional

Dreamscape The dream scene is set in: an elementary school located on the top of a slightly elevated hill. My home is far from the school, but not unreachable. On a pitch-black night, so dark you...

发布者 LuochuanAD，日期 2026-03-01

Token Cost Control in AIAgent

Cost Management

Background When building AI products, as the number of users grows, every LLM call and the prompt tokens used incur costs. So how can we minimize unnecessary expenses without affecting the resu...

发布者 LuochuanAD，日期 2026-02-28

User Intent Analysis of AIAgent

Router

Background User queries may involve simple information searches, tool invocations, casual chatting, and more. How can we perform intent analysis (or query classification) on user queries in the ...

发布者 LuochuanAD，日期 2026-02-23

Detailed Explanation and Reflection on the CAMEL Framework

CAMEL

Background CAMEL is the first large-model-based multi-agent framework designed with multiple role-playing agents that converse with each other to achieve a shared user task. Explanation of CA...

发布者 LuochuanAD，日期 2026-02-22

Building and Reflections on a Financial Statement Analysis AI Agent for ReAct RAG Agent

Architecture

Background Without complex RAG design, simply using LlamaIndex’s ReAct RAG Agent framework to analyze financial statements can achieve a level comparable to a junior business analyst. However, t...

发布者 LuochuanAD，日期 2026-02-21

Complete System Architecture of AIAgent

Architecture

Complete Architecture Structure 1. Intent Layer 2. Planning Layer 3. Execution Layer 4. Tool Layer 5. Memory Layer 6. Knowledge Layer 7. Policy & Guardrail Layer 8. Observability Layer 9. Persi...

发布者 LuochuanAD，日期 2026-02-19

Enterprise-level AI Agent for Router + ReAct + Function Calling

Architecture

Background User queries include both simple information retrieval and tool invocation, so a mature, reusable, and general system architecture is needed. Architecture User Query ↓ 1️⃣ Int...

发布者 LuochuanAD，日期 2026-02-18

Enterprise-level AI Agent for ReAct + Function Calling

ReAct Framework

Background In a complete AI Agent, tool invocation adopts the Function Calling concept from ChatGPT, connecting LLM and tool calls through defining a “Tools Schema.” Architecture User Query ...

发布者 LuochuanAD，日期 2026-02-17

Enterprise-level AI Agent for Plan-and-Execute + Function Calling

Plan-and-Execute Framework

Background In a complete AI Agent, the invocation of tools here follows the concept of ChatGPT’s Function Calling, connecting the LLM with tool calls by defining a “Tools Schema.” 1. Define t...

发布者 LuochuanAD，日期 2026-02-16

Plan-and-Execute Framework for Concepts and Applications

Plan-and-Solve

Background LangChain’s Plan-and-Execute framework is inspired by the Plan-and-Solve paper. Plan-and-Execute is well-suited for more complex long-term planning by breaking down complicated probl...

发布者 LuochuanAD，日期 2026-02-15

RAG Retrieval for Query ReWrite and Reranker

Query ReWriting

Background In designing the RAG system, after implementing Structural Chunks and Structural Prompt, the retrieval results achieved the desired accuracy. However, to further improve accuracy, du...

发布者 LuochuanAD，日期 2026-02-14

Detailed Explanation and Insights on the ReAct Framework

ReAct Prompt

Background Various papers propose multiple reasoning frameworks for intelligent agents (also called frameworks): CoT, ToT, LLM+P, etc. Among them, the ReAct framework is used as the reasoning e...

发布者 LuochuanAD，日期 2026-02-11

Structural Chunks Design of RAG

Structured Cutting

Introduction To improve the accuracy of retrieval and generation results in a RAG system, the core lies in the design of structured chunks. 1. Conventional Chunking Methods 1. Fixed-size Chu...

发布者 LuochuanAD，日期 2026-01-31

Explanation of the Prompt Framework and Structural Prompt

Structured Prompts

Introduction When developing different AI Agents, customized prompts are needed. Hence, professional, concise prompts that are easy for LLMs to understand become crucial. 1. Popular Prompt Fr...

发布者 LuochuanAD，日期 2026-01-11