Tag: 학습
All the articles with the tag "학습".
-
FEMS Project #3 — Streamlit Comparison Dashboard + 3-Backend Question Evaluation (Claude 9 / exaone 7 / gpt 7)
Put a Streamlit dashboard on top of FEMS RAG and threw the same question at 3 backends (exaone3.5:7.8b / claude-opus-4-8 / gpt-4o) simultaneously to compare. Q1 (air compressor anomaly in May) — exaone and gpt said 'no data', while Claude inferred 'weekend-hours anomaly' from summary stats alone without raw data + disclosed its limitations. Q2 (savings measures from a manager's perspective) — Claude 9/10 (incomplete due to token truncation), exaone 7/10 (broken index), gpt 7/10 (concise but hallucinated 'capacitor'). Results of 5-axis scoring.
-
MCP Study #1 — Getting Started with Model Context Protocol · First Call with stdio Server + Inspector
Starting to study MCP (Model Context Protocol) today. It's a standard protocol created by Anthropic that standardizes how LLMs access external systems. Comparing it with Tool Use → building my first MCP server (say_hello / add_numbers) with the Python SDK → verifying the stdio connection with MCP Inspector → building a fake energy management server that connects with FEMS. Connection failure due to venv python.exe path issue + fix.
-
The Full Flow of RAG Data Preparation — Selection, Cleansing, Chunking, Metadata, and Evaluation Sets
While working on the FEMS project, I got curious about 'what and how should go into a vector DB.' Key insight — data should be selected backward from 'questions that need to be retrieved,' not from the 'domain' criterion. Also covers cleansing / the effect of metadata (document title, section path) before chunks / OCR preprocessing for analog data / building an evaluation set — the full flow of the RAG data pipeline.
-
FEMS Project #2 — A Real 230-Page Corpus + Chunk Quality Gates + the Chinese-Character Mixing Trap in a Local LLM
Built a corpus of roughly 230 pages / 35,000 rows from the Korea BEMS Association guides, Korea Energy Agency materials, and the UCI Steel dataset. After paragraph-based chunking (target 800 chars) + bge-m3 + Chroma indexing, a quality gate (ratio of complete Hangul/ASCII characters) excluded 7 chunks from table-of-contents pages. Then I hit a trap — qwen2.5:7b mixed in Chinese characters on the second question and suffered generation collapse (spitting out unrelated Chinese city coordinates as GeoJSON). Partially fixed with temperature / system prompt → ultimately switched to exaone3.5:7.8b for clean handling.