cfgnotes https://cfgnotes.com/ en All rights reserved 2026, cfgnotes Thu, 28 May 2026 10:00:00 GMT Hexo Qwen3.5 Pitfalls: Thinking Chain and flash-attn https://cfgnotes.com/lora-elderly/qwen35-pitfalls/ Verification and vLLM must-disable thinking, flash-linear-attention warnings, MPS+PEFT device strategies. cfgnotes Thu, 28 May 2026 10:00:00 GMT vLLM Deployment: Dynamic Mounting of LoRA and OpenAI-Compatible API https://cfgnotes.com/lora-elderly/vllm-deployment/ vllm serve --enable-lora, curl/Python invocation, production checklist (excluding weight merging). cfgnotes Thu, 28 May 2026 10:00:00 GMT LoRA Effect Validation: The Principle of verify_lora.py and Practical Tests on Mac https://cfgnotes.com/lora-elderly/verify-lora/ Comparison before and after fine-tuning, enable_thinking=False, MPS inference and companion-style rubric. cfgnotes Wed, 27 May 2026 16:00:00 GMT Understanding the LoRA Training Curve: 750-Step Real Trading Review https://cfgnotes.com/lora-elderly/training-metrics/ Interpretation of four milestones in all_logs.log, difference between train_loss and stepwise loss, method for judgment when there is no eval set. cfgnotes Wed, 27 May 2026 14:00:00 GMT Single Card SFT Practice (Part 2): SFTTrainer and Training Loop https://cfgnotes.com/lora-elderly/sft-training-part2/ SFTConfig parameters, paged_adamw_8bit, Trainer injection of LoRA, final_lora saving and checkpoint. cfgnotes Wed, 27 May 2026 12:00:00 GMT Single-Card SFT Practical (Part 1): Tokenizer, Base Model, and LoRA Configuration https://cfgnotes.com/lora-elderly/sft-training-part1/ train_lora_single.py Steps 1–4: device_map, LoraConfig, apply_chat_template training mode. cfgnotes Wed, 27 May 2026 10:00:00 GMT Qwen3.5-4B LoRA Fine-tuning Environment Setup and Model Preparation https://cfgnotes.com/lora-elderly/environment-setup/ PyTorch/TRL/PEFT version pinning, ModelScope download, V100 training and Mac MPS validation division of labor. cfgnotes Tue, 26 May 2026 16:00:00 GMT LoRA Principle: Training Only 0.25% of Parameters https://cfgnotes.com/lora-elderly/lora-principles/ Low-rank decomposition, meaning of r/alpha/target_modules, comparison of adapter_config.json and SFTTrainer injection points. cfgnotes Tue, 26 May 2026 14:00:00 GMT Training Set Design: 1000 JSONL and Elderly Psychological Model https://cfgnotes.com/lora-elderly/dataset-design/ messages three-role format, five categories of psychological themes, system prompt design, and load_jsonl_data consumption method. cfgnotes Tue, 26 May 2026 12:00:00 GMT pydantic-settings Configuration Management: Best Practices for API Key Masking and Multi-Environment Configuration https://cfgnotes.com/ican/pydantic-settings/ BaseSettings, .env priority, llm_config_dict desensitization, and module-level singleton. cfgnotes Tue, 26 May 2026 10:00:00 GMT Why Use LoRA for Elderly Emotional Companionship AI https://cfgnotes.com/lora-elderly/why-lora-elderly-companion/ Vertical scenario selection: empathetic companionship vs. general preaching, LoRA SFT pipeline overview and V100 actual test 41-minute training data. cfgnotes Tue, 26 May 2026 10:00:00 GMT Intelligent Dialogue Engine in Practice — Multi-turn Context Management and Smart Chart Recommendation https://cfgnotes.com/case/data-analysis/chat-engine/ chat_engine linear pipeline: session persistence, multi-turn context, rule engine chart recommendation and report generation, non-LLM chart selection. cfgnotes Tue, 26 May 2026 10:00:00 GMT Implementing a Frontend Drag-and-Drop Interaction System —— Drag API + Grid + mousedown https://cfgnotes.com/case/data-analysis/drag-drop-editor/ DashboardBuilder: Component library HTML5 drag-and-drop, canvas mousedown move and resize, grid snapping and layout persistence. cfgnotes Tue, 26 May 2026 00:00:00 GMT Data Dashboard Two-Stage Separation Architecture: Design Time vs Runtime Decoupling https://cfgnotes.com/case/data-analysis/dashboard-two-phase/ At design time, resolve writes to query_config.sql; at runtime, render only executes SQL; large-screen refresh at 200ms level, zero LLM cost. cfgnotes Tue, 26 May 2026 00:00:00 GMT Production Deployment and Performance Optimization — Asynchronous Architecture, Caching Strategy, and Monitoring https://cfgnotes.com/case/data-analysis/production-deployment/ Distinguish between the current FastAPI monolith (uvicorn + Vite) and the planned Redis, Compose, Prometheus/ELK; local run-through and evolution roadmap. cfgnotes Tue, 26 May 2026 00:00:00 GMT FastAPI Middleware Design: Custom RateLimitMiddleware and CORS Configuration Practice https://cfgnotes.com/ican/fastapi-middleware/ Sliding window rate limiting, /health exemption and CORSMiddleware MVP configuration. cfgnotes Mon, 25 May 2026 16:00:00 GMT Full Record of Pitfalls in Docker Deployment of AI Projects: PyTorch CPU Version + Chinese Fonts + Frontend Build https://cfgnotes.com/ican/docker-deploy/ Multi-stage Dockerfile, CPU version PyTorch, fonts-noto-cjk, and node_modules overwriting pitfalls. cfgnotes Mon, 25 May 2026 14:00:00 GMT Prompt Strategy for Conversational AI Agent: 5-Stage Progressive Information Mining https://cfgnotes.com/ican/guide-prompt/ GUIDE_SYSTEM_PROMPT, five-stage template, and check_sufficiency dual-track stopping criteria. cfgnotes Mon, 25 May 2026 12:00:00 GMT NL→SQL Translation Engine — Semantic Model Injection and Security Verification https://cfgnotes.com/case/data-analysis/nl2sql-engine/ Semantic model injection, LiteLLM generating SQL, SELECT validation and blacklist filtering; a complete pipeline for interfacing dialogue and large-screen resolve. cfgnotes Mon, 25 May 2026 10:00:00 GMT RIASEC Holland Assessment Prompt Engineering: Let AI Scientifically Assess Career Interests https://cfgnotes.com/ican/riasec-prompt/ Analyze RIASEC prompt structure, Holland Code output, and CareerMatcher integration. cfgnotes Mon, 25 May 2026 10:00:00 GMT Prompt Design for AI Agent: How to Make LLM Stably Output Structured JSON https://cfgnotes.com/ican/prompt-json/ json_object mode, parse_json_from_text four-level degradation, and ResumeParser Prompt structure. cfgnotes Sun, 24 May 2026 16:00:00 GMT Python PDF Generation in Practice: ReportLab + matplotlib for Professional Chinese Reports https://cfgnotes.com/ican/pdf-report/ Direct parsing of Markdown with ReportLab, CJK fonts, and embedding matplotlib charts into PDF. cfgnotes Sun, 24 May 2026 14:00:00 GMT SQLAlchemy Automatic Migration Solution: Technique to Add Fields Automatically Without Alembic https://cfgnotes.com/ican/sqlalchemy-auto-migrate/ _auto_migrate: add columns at startup, FastAPI lifespan integration, and MVP security strategy. cfgnotes Sun, 24 May 2026 12:00:00 GMT FastAPI SSE Streaming Output + WebSocket Progress Push: Real-Time Communication Solution for AI Applications https://cfgnotes.com/ican/sse-websocket/ chat/stream SSE, ws_manager reports progress, and run_analysis_pipeline asynchronous push. cfgnotes Sun, 24 May 2026 10:00:00 GMT Metadata Intelligent Management System —— LLM-driven Semantic Model Construction https://cfgnotes.com/case/data-analysis/metadata-management/ LLM automatically parses Chinese field names and table relationships, assembles a semantic model in real-time and injects it into NL→SQL; explains the difference from planning capabilities such as Redis caching. cfgnotes Sun, 24 May 2026 10:00:00 GMT FastAPI + LangChain Integration Best Practices: Unified LLM Call Interface Design https://cfgnotes.com/ican/llm-integration/ providers.py dual model, invoke_llm_with_json and ChatOpenAI compatible API switching. cfgnotes Sat, 23 May 2026 16:00:00 GMT LangGraph Error Handling and Fault Tolerance Design: 5 Strategies to Prevent AI Agent System Crashes https://cfgnotes.com/ican/error-handling/ Health check, node try/except, JSON four-layer degradation, and template degradation when Ollama is unavailable. cfgnotes Sat, 23 May 2026 14:00:00 GMT LangGraph Double-Layer StateGraph Nesting: Architectural Practice for Reducing Agent Coupling https://cfgnotes.com/ican/nested-stategraph/ Double-layer nesting of create_workflow and create_guide_graph, subgraph ainvoke and decoupling test. cfgnotes Sat, 23 May 2026 12:00:00 GMT LLM Unified Gateway Design — LiteLLM Abstraction, Prompt Engineering, and Response Cleaning https://cfgnotes.com/case/data-analysis/llm-gateway/ Using LiteLLM to unify DeepSeek/OpenAI calls, embedding Prompt and multi-layer response cleaning, achieving zero-cost model switching and stable JSON/SQL output. cfgnotes Sat, 23 May 2026 10:00:00 GMT LangGraph Conditional Routing in Practice: Implementing Agent Loop Dialogue and Forced Exit Mechanism https://cfgnotes.com/ican/conditional-routing/ route_after_guide, should_continue, and loop limit, Guide loop and forced handoff. cfgnotes Sat, 23 May 2026 10:00:00 GMT