Qwen3.5 Pitfalls: Thinking Chain and flash-attn
0. Series Closure This Article Position Upstream This Article Output Downstream Article 9/10 Article 08 Inference Verification Qwen3.5 Specific Issue...
Read more →cfgnotes
Agent · RAG · Engineering Deep Dives
From Known to Unknown — Practical guides for building production-grade AI systems.
Browse Topics ↓Latest Articles
vLLM Deployment: Dynamic Mounting of LoRA and OpenAI-Compatible API
0. Series Closure This Article Upstream Output Downstream Part 10/10 Training + Validation Production-ready HTTP API Business App / Mini Program...
Read more →LoRA Effect Validation: The Principle of verify_lora.py and Practical Tests on Mac
0. Series Loop Position Upstream Output Downstream Post 8/10 Post 07: Loss Convergence Qualitative conclusion: Whether it’s “gentler, less preachy” Po...
Read more →Understanding the LoRA Training Curve: 750-Step Real Trading Review
0. Series Closure Position in Series Upstream This Article Output Downstream Article 7/10 Article 06: Training Complete Metric interpretation, determi...
Read more →Single Card SFT Practice (Part 2): SFTTrainer and Training Loop
0. Series Overview This Article Upstream Output Downstream Article 6/10 Article 05 – Dataset Ready final_lora/, 15 checkpoints Article 07 – Reading Lo...
Read more →Single-Card SFT Practical (Part 1): Tokenizer, Base Model, and LoRA Configuration
0. Series Completion Position in Series Upstream This Output Downstream Article 5/10 Article 04: Environment Ready Tokenizer, Base Model, LoraConfig,...
Read more →Qwen3.5-4B LoRA Fine-tuning Environment Setup and Model Preparation
0. Series Closing the Loop This Article’s Position Upstream This Article’s Deliverable Downstream Article 4/10 Article 03: Principles Runnable GPU env...
Read more →LoRA Principle: Training Only 0.25% of Parameters
0. Series Closed Loop Position in Series Upstream This Post’s Output Downstream Post 3/10 Post 02: Data into Model Understand r/alpha/target...
Read more →Training Set Design: 1000 JSONL and Elderly Psychological Model
0. Series Closure Position in Series Upstream This Article’s Output Downstream Article 2/10 Article 01: Scenario Definition Trainable JSONL Specificat...
Read more →pydantic-settings Configuration Management: Best Practices for API Key Masking and Multi-Environment Configuration
0. Series Loop (Follow Along Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analysis_p...
Read more →Why Use LoRA for Elderly Emotional Companionship AI
0. Series Overview Position in Series Upstream Output Downstream Article 1/10 Business scenario definition Technical selection conclusion, project bou...
Read more →Intelligent Dialogue Engine in Practice — Multi-turn Context Management and Smart Chart Recommendation
🎯 Core Questions of This ChapterHow to build a truly “intelligent” conversational system? Problems with ordinary chatbots: ❌ No memory: Each conversation roun...
Read more →Implementing a Frontend Drag-and-Drop Interaction System —— Drag API + Grid + mousedown
🎯 Core Questions of This ChapterHow to achieve a smooth and intuitive drag-and-drop interaction experience? Challenge Pain Points of Traditional Solutions O...
Read more →Data Dashboard Two-Stage Separation Architecture: Design Time vs Runtime Decoupling
🎯 Core Question of This ChapterWhat is the core contradiction of Dashboards? Dimension Problems with Traditional Approaches Our Solution Performance LLM...
Read more →Production Deployment and Performance Optimization — Asynchronous Architecture, Caching Strategy, and Monitoring
🎯 Core Chapter QuestionsHow to safely and efficiently deploy a development environment application to production? Challenge Traditional Pain Points Our Solu...
Read more →FastAPI Middleware Design: Custom RateLimitMiddleware and CORS Configuration Practice
0. Series Loop (Follow Along Even Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analy...
Read more →Full Record of Pitfalls in Docker Deployment of AI Projects: PyTorch CPU Version + Chinese Fonts + Frontend Build
0. Series Loop (Follow along without public source code)End-to-end pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_p...
Read more →Prompt Strategy for Conversational AI Agent: 5-Stage Progressive Information Mining
0. Series Loop (Follow Without Publishing Source)End-to-end pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_pipeline...
Read more →NL→SQL Translation Engine — Semantic Model Injection and Security Verification
🎯 Core Questions of This ChapterHow to make LLM generate accurate, safe, and executable SQL? This is the core challenge of the entire system: ❌ The LLM doesn’...
Read more →RIASEC Holland Assessment Prompt Engineering: Let AI Scientifically Assess Career Interests
0. Series Loop (Follow Along Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analysis_p...
Read more →Prompt Design for AI Agent: How to Make LLM Stably Output Structured JSON
0. Series Loop (Follow Along Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analysis_p...
Read more →Python PDF Generation in Practice: ReportLab + matplotlib for Professional Chinese Reports
0. Series Loop (Follow along without open-source code)End-to-end pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_pip...
Read more →SQLAlchemy Automatic Migration Solution: Technique to Add Fields Automatically Without Alembic
0. Series Loop (Read Along Without Public Source Code)End-to-End Pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_pip...
Read more →FastAPI SSE Streaming Output + WebSocket Progress Push: Real-Time Communication Solution for AI Applications
0. Series Loop (Follow Along Without Public Source Code)End-to-End Pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_p...
Read more →Metadata Intelligent Management System —— LLM-driven Semantic Model Construction
🎯 Core Questions of This ChapterHow does an LLM understand your database structure? This is the core challenge of all NL→SQL systems: ❌ The LLM doesn’t know w...
Read more →FastAPI + LangChain Integration Best Practices: Unified LLM Call Interface Design
0. Series Loop (Readalong Without Open Source)End-to-End Pipeline: Vue frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analysis_pipeline (P...
Read more →LangGraph Error Handling and Fault Tolerance Design: 5 Strategies to Prevent AI Agent System Crashes
0. Series Loop (Follow Along Without Open Source Code)End-to-End Chain: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_pipeli...
Read more →LangGraph Double-Layer StateGraph Nesting: Architectural Practice for Reducing Agent Coupling
0. Series Loop (Read Along Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analysis_pip...
Read more →LLM Unified Gateway Design — LiteLLM Abstraction, Prompt Engineering, and Response Cleaning
🎯 Core Chapter QuestionsIn AI application development, how to elegantly call LLM APIs? Hard-coding with requests.post()? Or writing a set of calling logic for...
Read more →LangGraph Conditional Routing in Practice: Implementing Agent Loop Dialogue and Forced Exit Mechanism
0. Series Loop (Read Along Without Public Source Code)End-to-End Pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_pip...
Read more →LangGraph State Management Detailed: Correct Usage of TypedDict + Annotated Reducer
0. Series Loop (Readable Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi‑turn SSE → run_analysis_pipel...
Read more →Holland RIASEC + OpenAI-Compatible API: AI Career Assessment Engineering Implementation
0. Series Closed Loop (Read Along Even Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_...
Read more →LangGraph Multi-Agent Orchestration Practice: How 5 Agents Collaborate to Complete Career Planning Analysis
0. Series Loop (Follow Along Without Public Source Code)End-to-end pipeline: Vue frontend → api/routes/chat.py → Guide multi-turn SSE → run_analysis_pipeline (p...
Read more →NLP MySQL Intelligent Data Analysis Platform — Project Overview and Technology Selection
📖 Introduction: Why This Project?In daily enterprise operations, data querying is one of the most frequent needs, but the reality is harsh: Business users: Wa...
Read more →Building an AI Agent System from Scratch: FastAPI + LangGraph Smart Career Planning Platform
0. Series Loop (Read Along Without Public Source Code)End-to-End Pipeline: Vue Frontend → api/routes/chat.py → Guide Multi-turn SSE → run_analysis_pip...
Read more →Bid Farewell to Retrieval Hallucinations! Step-by-Step Guide to Building an Enterprise-Grade RAG Data Pipeline (with One-Click Docker Deployment)
Pain Point: Using Generic Embedding Models for RAG, Getting Irrelevant Answers in Vertical Domains?Have you ever encountered this: Using OpenAI’s text-embeddin...
Read more →Docker Containerized Deployment: Solving Timezone Mismatch and Log Mounting Issues
1. IntroductionWhen deploying with containers, it is common to encounter issues where the time inside the container differs from the host (typically by 8 hours)...
Read more →JWT Authentication: AccessToken has expired, how should RefreshToken be used?
1. IntroductionIn JWT (JSON Web Token) based authentication systems, AccessToken is typically set with a short lifetime (15 minutes to 1 hour) to reduce securit...
Read more →Hand-build a Simplest Permission Center: RBAC Model Design and Implementation
1. IntroductionIn a microservices and front-end/back-end separation architecture, permission management is one of the fundamental pieces of system security...
Read more →Introduction to Hybrid Development: A Step-by-Step Guide to Writing a JSBridge
1 IntroductionHybrid mobile development has become the mainstream choice for balancing multi-platform development efficiency with native experience. While WebVi...
Read more →Frontend Solution for Embedding Agent on Web
1. IntroductionAI Agents are evolving from standalone applications to embedded interactive experiences on the web. The core challenge for frontend teams is: how...
Read more →Agent Request Timeout and Slow Response Optimization
1. IntroductionThis article introduces the issues of request timeout and slow responses in Agent systems caused by network fluctuations, long tool execution tim...
Read more →Agent Integration with Local Knowledge Base for RAG Combined Use
1. IntroductionTraditional RAG lacks proactive reasoning and multi-step retrieval capabilities in complex Q&A scenarios. This article introduces how to inte...
Read more →Agent Task Auto-Decomposition Logic Development
1. IntroductionAs large model-driven agents are gradually deployed in enterprise production environments, a common problem emerges: With the same underlying lar...
Read more →PDF extraction always missing tables? PyMuPDF + PaddleOCR-VL hybrid solution in practice (with MLX acceleration)
Pain point: Why is your table always empty when extracting PDF with pdfplumber?Let’s start with a real scenario: You have a 50-page PDF product technical specif...
Read more →Implementation of A2A Multi-Agent Inter-Communication and Invocation
1. IntroductionAs AI Agent application scenarios evolve from single-point tasks to complex workflows, multi-agent cross-platform collaboration has become a nece...
Read more →Implementation of A2A Multi-Agent Inter-Communication and Invocation
1. IntroductionAs AI Agent application scenarios evolve from single-point tasks to complex workflows, multi-agent cross-platform collaboration has become a nece...
Read more →Rapid Adaptation of MCP Protocol Integration with Agent
Quick Adaptation for MCP Protocol Access in Agents1. IntroductionIn LLM application development, a common problem is the need to write adaptation code for each...
Read more →Avoiding Pitfalls in Function Call Debugging
1. IntroductionFunction Call is a core capability that enables large language models to interact with external systems, allowing the model to trigger the execut...
Read more →Agent Custom Tool Development and Integration Methods
Developing and Integrating Custom Tools for Agents1. IntroductionAgents extend their capabilities through tool calls (Function Call), which is currently the mai...
Read more →Agent Development and Deployment: Building a Production-Grade Intelligent System from Scratch
1. Introduction: Easy to Develop Agents, Hard to Deploy ThemThis article will focus on best practices for Agent development and deployment, systematically break...
Read more →How to Chunk RAG Without Losing Context? 5 Strategies from Beginner to Production-Grade (With Decision Tree for Selection)
Pain Point: Should chunk_size be 500 or 1000? Why is the effect still poor after adjusting N times?Chunking is the most underestimated step in a RAG system. Man...
Read more →Agent Long-Term Persistent Memory Practical Setup
Practical Guide to Building Long-Term Persistent Memory for Agents - Internal Knowledge Base Document1. IntroductionIn real-world agent deployments, stateless d...
Read more →Agent Short-term Session Memory Implementation Plan
1. IntroductionThe token limit of large model context windows (typically 4K–128K tokens) determines the amount of information that can be carried in a single co...
Read more →Agent Character Persona and Task Logic Design Method
1. IntroductionAs LLM-driven agents are gradually deployed in enterprise production environments, a common issue has emerged: given the same underlying model, w...
Read more →BGE-M3 Local Fine-Tuning Practice: From Scratch to Production-Level Deployment (with Complete Code)
📊 Table of Contents Why Fine-tune BGE-M3? Core Capabilities of BGE-M3 Fine-tuning Environment Setup Guide Data Preparation: Building a High-Quality Training Se...
Read more →Differences in Selection Between Lightweight Agent and Enterprise Heavy-Duty Agent
IntroductionUnder the dual trends of edge computing and enterprise digitalization, the technical selection of Agents involves a trade-off between efficiency and...
Read more →Core Essential Capabilities of Autonomous Decision-Making Agents
1. IntroductionCurrent AI systems in production environments often face scenarios that are dynamic, complex, and require independent operation. This demands tha...
Read more →Difference between Agent and ordinary large model chatbot
The Essential Difference Between Agent and Ordinary Large Model Conversational BotIntroductionWith the widespread application of Large Language Models (LLMs), c...
Read more →Milvus Production Environment Collection Design + HNSW Tuning Practical Guide
📊 Table of Contents Why Meticulously Design a Milvus Collection? Collection Schema Best Practices Deep Dive into HNSW Index Principles Complete HNSW Para...
Read more →A Plain Explanation of What Is an AI Intelligent Agent
Understanding AI Agent: A Simple ExplanationIntroductionLarge Language Models (LLMs) have demonstrated impressive capabilities — they can answer questions, writ...
Read more →RAG Implementation: Production Environment Deployment and Performance Monitoring Practice
1. IntroductionMoving a RAG system from prototype to production typically encounters three core challenges: uncontrollable response latency, fluctuating retriev...
Read more →Table 4-level Vectorization Scheme: Let RAG System Truly Understand Structured Data
📊 Table of Contents Why Tables are a Nightmare for RAG Systems? Core Idea of 4-Level Table Vectorization Detailed Design and Scenarios of 4 Levels Complete Cod...
Read more →RRF Multi-channel Fusion Ranking: The Secret Weapon That Improves RAG Retrieval Accuracy by 30%+
📊 Table of Contents Why Do We Need Multi-Channel Retrieval Fusion? Deep Dive into RRF Algorithm Core Principles Multi-Channel Retrieval Architecture in RAG Sys...
Read more →MySQL+Milvus+MinIO Three-Storage Dual-Write Architecture: Building an Enterprise-Level RAG Data Foundation
📊 Table of Contents Why Does a RAG System Need a Triple‑Storage Architecture? Triple‑Storage Responsibility Division and Design Philosophy Dual‑Write Consisten...
Read more →RAG Evaluation: End-to-End Metric Design and Effect Evaluation Framework
1. Introduction: Why RAG Evaluation Can’t Rely on “Feelings” Alone?Imagine this scenario: You painstakingly build a RAG system. The knowledge base contains tens...
Read more →RAG Advanced: Agentic RAG — Dynamic Tool Calling and Iterative Optimization
1. Introduction: The Limitations of Traditional RAG and the Value Proposition of Agentic RAGDo you still remember the excitement of deploying your first RAG sys...
Read more →RAG Advanced: Multimodal RAG — Image-Text Mixed Retrieval and Generation
1. Introduction: The “Blindness” Dilemma of Traditional RAG and the Multimodal BreakthroughHave you ever encountered this scenario: you feed a PDF report full o...
Read more →RAG Online Part: Generation Optimization — Self-RAG and Adaptive Retrieval
1. Introduction: The Bottleneck in Generation – Why Self-RAG and Adaptive Retrieval?RAG (Retrieval-Augmented Generation) has a common pain point: all queries ar...
Read more →RAG Online Part: Retrieval Optimization — Multi-Channel Recall and Result Fusion
1. Introduction: Single Path Not Enough – How Multi-Path Recall Solves the “Lopsided” Retrieval Problem?After deploying an RAG (Retrieval-Augmented Generation)...
Read more →RAG Online Part: Retrieval Optimization — HyDE and Query Expansion Techniques
Introduction: Why is your RAG retrieval always inaccurate? Starting with HyDEImagine this scenario: You’ve built an enterprise knowledge base RAG (Retrieval-Aug...
Read more →RAG Offline Part: Embedding Model Selection and Domain Adaptation Fine-tuning
1. Introduction: Why Is the Embedding Model the “Invisible Bottleneck” of RAG?Have you ever encountered this scenario? Your RAG (Retrieval-Augmented Generation)...
Read more →RAG Offline Part: Metadata Enhancement and Knowledge Graph Fusion Preprocessing
1. Introduction: Why Does RAG Offline Preprocessing Need Metadata Enhancement and Knowledge Graphs?Hi, I’m a tech blogger. Today, let’s talk about a critical ye...
Read more →RAG Offline Part: Multi-Source Heterogeneous Data Cleaning and Deduplication Strategy
Introduction: When “Multi-Source Data” Becomes the Nightmare and Turning Point of RAGImagine this scenario: you’re developing a smart Q&A system for e-comme...
Read more →