We continuously share large‑model know‑how, helping you master core skills—LLM, RAG, fine‑tuning, deployment—from zero to job offer, tailored for career‑switchers, autumn recruiters, and those seeking stable large‑model positions.
This article explains how misplacing retry_count in LangGraph's user-level Store instead of thread-level State causes cross-task contamination, detailing the distinct lifecycles of State, Runtime Context, and Store with code examples and engineering checks.
The article exposes a flaw in AI agent evaluation where a refund issued before attachment verification still receives a perfect score because the evaluator only checks if required tools were eventually called, not whether preconditions were satisfied before write operations.
An AI agent correctly retrieves a 'no refund' policy but still executes a $50 refund, exposing the critical gap between policy retrieval and runtime enforcement; the article argues authorization must be enforced at the tool gateway with bound decisions, not just in model context, and outlines regression tests for deny paths.
This article explains why embedding model selection must precede reranker evaluation in RAG systems, detailing a three-stage evaluation methodology using business-specific data to measure recall, ranking quality, and end-to-end answer validity, while accounting for engineering constraints like latency, resource usage, and failure modes.
The article shows that merely limiting an Agent to a fixed number of steps (e.g., max_steps=8) cannot guarantee progress, and proposes a detailed trace‑based contract, fingerprinting, and soft‑stop criteria to reliably identify and halt loops where no new evidence is produced.
The article explains that asyncio.gather only propagates the first exception without cancelling other awaitables, distinguishes fan‑out/fan‑in from a Supervisor role, and demonstrates a pure‑Python pattern for classifying required versus optional branches, handling failures, and merging results safely.
The article explains how using -100 as an ignored label in PyTorch cross‑entropy loss silently masks non‑assistant tokens, how to locate assistant spans via prefix‑difference, the trade‑offs between supervising only the final reply versus all assistant turns, and the essential pre‑training checks to avoid hidden masking errors in multi‑round SFT.
The article shows that a retrieval‑augmented generation system can pass a similarity threshold yet give a wrong answer because the retrieved evidence does not cover all required conditions, and it proposes a condition‑checklist approach instead of relying on raw scores.
The article analyzes a deterministic rule called false_promise that flags agent replies claiming completed actions without corresponding tool calls, explains how tense affects verification, proposes a four‑step “claim‑check” process, and shows how scoring caps and regression samples expose both true violations and false‑positive edge cases.
The article shows that relying on filenames for deduplication creates both duplicate entries and missed updates, explains why a two‑level identity (stable document ID and chunk ID) is required, and presents practical tests and design guidelines for reliable incremental knowledge‑base ingestion.
