This aggregates CodeNodeX AI community articles and discussions related to 大模型, covering concept explanations, tool comparisons, engineering practices, common issues, and indie hacking experience.
If you're researching 大模型, start with high-quality long-form articles, then join post discussions for a complete understanding-to-practice path.
A prompt is not a "wish" but a specification for the model. This article explains the six elements of a structured prompt, advanced techniques such as CoT, few-shot examples, and XML segmentation, defensive phrasing, and a reusable debugging method that turns intuition-driven prompt tuning into an engineering process.
A high-quality prompt is not a magic instruction but an engineering system comprising a task contract, context, examples, an output structure, and an evaluation set. This article presents a reusable prompt design method and explains how version control and regression testing enable stable iteration.
"Model leaderboards cannot replace business evaluation. This article designs a five-layer LLM application evaluation framework spanning evaluation-set construction, rule-based metrics, model judges, human review, and production observation to support model selection, Prompt regression, and RAG optimization."
Agent tasks for research, procurement, and batch data processing often span minutes or even days. This article explains how to persist task snapshots, restore checkpoints, handle timeouts, and avoid duplicating external side effects after restarts.
As memory grows, you can't save tokens by overwriting old summaries. This article designs fixed input snapshots, derived lineage, key-fact fidelity tests, invalidation propagation, and reversible switching.
User preferences, project decisions, and roles all change. This article separates facts using valid time and recorded time, explicitly handles supersession, late-arriving, and concurrent conflicts, and avoids last-writer-wins producing incorrect memory.
Preferences and facts a model extracts from a conversation cannot become long-term memory directly. This article designs candidate proposals, deterministic validation, deduplicated commits, revocation lineage, and a contamination-rate gate.
Passing offline evaluation does not mean an agent is reliable in production. This article designs online success rate, human takeover, rejection quality, tool recovery, and cost metrics to help teams use layered observation to judge whether a version deserves more traffic.
Multi-step agent errors often occur in intermediate tool results and state changes. This article explains how to save desensitized trajectories, build deterministic replays, handle external dependencies, and turn production incidents into reliable regression tests.
A production agent cannot be just a while loop. This article uses an explicit state machine to split planning, tool calls, waiting for approval, recovery, and terminal states, and provides transition constraints, persistence fields, and testing methods.
Neither bot visits nor traditional impressions directly prove a page is cited in AI answers. This article establishes a repeatable question panel, citation normalization, and intervention records to avoid manufacturing GEO conclusions from scattered screenshots.
For RAG and semantic search, choosing a vector database is unavoidable. This article compares pgvector, Qdrant, Milvus, Chroma, and Pinecone by data scale, latency, operational cost, and hybrid search, with clear recommendations for when to use each.