Sanguo RAG Resolution Loop

這個 skill 用來把 observed mentions -> unresolved labels -> MCQ 裁決 -> decisions/manual roster -> rebuild/rescan 做成可重複循環。

Unity 對照：這類似一個資料版 AssetPostprocessor。腳本負責 deterministic import/rebuild，人工或研究 agent 只裁決少數不確定 label。

When to Use

• 使用者提到 unresolved、alias review、observed mentions、manual roster seeds、文本稱呼表、正式對照表。
• 使用者想「繼續跑循環」、「產生選擇題」、「套用 A/B/C/D 裁決」、「收斂三國 RAG 人名解析」。
• 使用者想把候選名詞交給 agent 查證是否為三國人物、地名、官稱或切詞噪音。
• 同一批名詞會反覆去固定網站或百科來源查證，想先壓成 compact term lookup CLI 再交給 reviewer。

Core Files

• Loop script: server/npc-brain/pipelines/sanguo-rag/run_resolution_loop.py
• Clean/split script: server/npc-brain/pipelines/sanguo-rag/clean_and_split.py
• Apply answers: server/npc-brain/pipelines/sanguo-rag/apply_triage_answers.py
• Research brief: server/npc-brain/pipelines/sanguo-rag/generate_term_research_brief.py
• Decisions: server/npc-brain/pipelines/sanguo-rag/config/unresolved-triage-decisions.json
• Manual seeds: server/npc-brain/pipelines/sanguo-rag/config/manual-roster-seeds.json
• MCQ output: artifacts/data-pipeline/sanguo-rag/extracted/resolution-loop/unresolved-triage-choices.md
• Answers template: artifacts/data-pipeline/sanguo-rag/extracted/resolution-loop/unresolved-triage-answers.todo.json

Pilot ETL Chunk Learning Mode

當使用者想學「LangChain text splitters / LangGraph / LangFlow 到底該放在哪」，先不要把整條 unresolved loop 改成框架驅動；先用 張飛長坂橋 這種單章小樣本做對照實驗。

Step 1. 保留 deterministic baseline

先跑原本的 clean_and_split.py，保留章回與 paragraph index：

python server/npc-brain/pipelines/sanguo-rag/clean_and_split.py \
  --input <source.md> \
  --output-root artifacts/data-pipeline/sanguo-rag/markdown \
  --overwrite

這會得到：

• source.md
• chapters/ch_###.md
• chapters-manifest.json

這一層是 deterministic evidence layer，類似 Unity 先把 raw asset import 成可追蹤 metadata。

Step 2. 加上 LangChain text splitters 做 chunk 對照

若要比較 paragraph-based 切法與 LLM 常用 chunking，再跑：

python server/npc-brain/pipelines/sanguo-rag/clean_and_split.py \
  --input <source.md> \
  --output-root artifacts/data-pipeline/sanguo-rag/markdown \
  --chunk-with-langchain \
  --chunk-size 500 \
  --chunk-overlap 80 \
  --overwrite

額外輸出：

• chunks/<chapter_id>/<chunk_id>.md
• chunks-manifest.json

學習重點：

• chapters-manifest.json 看 paragraph index / source offset。
• chunks-manifest.json 看 chunk size / overlap / source refs。
• 比較哪種切法在 E-5a 對話消歧、E-5b 事件抽取、E-6 keyword pack 比較穩。

Step 3. LangGraph 放在有分支狀態時才值得

LangGraph 不適合拿來取代 deterministic preprocessing。它最有價值的地方是：

• chunk -> extract -> validate -> retry -> review 這種有狀態、有分流的 LLM pipeline
• high-confidence -> publish / low-confidence -> review queue 這種圖狀流程
• 抽取失敗後需要保留 intermediate state 與重試策略的情境

若目前只是單章 PoC，先不用急著上 LangGraph。

Step 4. LangFlow 主要用來學習與展示，不是正式 ETL 依賴

LangFlow 最適合：

• 視覺化展示 prompt / parser / retriever 怎麼串
• demo 時快速調參
• 教學用地看 chain 每一步輸入輸出

但正式 pipeline 還是應以 Python 腳本為主，避免把核心 ETL 綁死在視覺化編排工具。

建議判斷

• 只做清洗、拆章、保留 paragraph index：先用純 Python + Pydantic。
• 要比較 chunk size / overlap 對抽取效果的影響：加 LangChain text splitters。
• 要做多步驟抽取、驗證、重試、分流：再考慮 LangGraph。
• 要做教學 demo 或視覺化展示：再補 LangFlow。

Standard Procedure

1. Pre-flight: read docs/keep.summary.md and use Traditional Chinese.
2. If text files will be edited, load encoding-touched-guard and run touched encoding checks before finishing.
3. Run one loop:

python server/npc-brain/pipelines/sanguo-rag/run_resolution_loop.py --top 30

4. Show the user the generated MCQ path and summarize the first few labels.
5. If the user filled answers, apply them:

python server/npc-brain/pipelines/sanguo-rag/apply_triage_answers.py
python server/npc-brain/pipelines/sanguo-rag/run_resolution_loop.py --top 30

6. Verify:

python -m py_compile \
  server/npc-brain/pipelines/sanguo-rag/collect_observed_mentions.py \
  server/npc-brain/pipelines/sanguo-rag/run_resolution_loop.py \
  server/npc-brain/pipelines/sanguo-rag/apply_triage_answers.py \
  server/npc-brain/pipelines/sanguo-rag/generate_term_research_brief.py

7. Encoding check touched files:

npm run check:encoding:touched -- --files <changed-text-files>

Decision Semantics

• A person: a real person label. Requires personRecord.generalId and personRecord.faction; apply_triage_answers.py adds it to manual roster seed.
• B noise: not a person. Adds to noiseLabels; collector emits it as excluded, not unresolved.
• C ambiguous: still needs review or is useful but not resolvable now. Adds to ambiguousLabels; collector emits it as review-pending.
• D defer: no action; it stays unresolved and will appear again.

Do not force all unresolved labels into person seeds. The goal is zero unclassified labels, not zero non-person labels.

Web Research Delegation

若同一批名詞會反覆查固定來源，不要每次都直接委派瀏覽器研究。先判斷是否該用 agent-cli-factory 建一層 repo-local term lookup CLI，例如 3klife-sanguo-term-lookup，把輸出壓成 --compact --json 的來源命中摘要、候選 person/noise 判斷與 citation 線索。

CLI-first 的目標不是跳過人工審核，而是先把「明顯不是人名」或「明顯命中既有來源」的候選縮成小結果集，讓 web-capable agent 只處理真正模糊的標籤。

現成 term lookup CLI：

node tools_node/agent-clis/3klife-sanguo-term-lookup.js \
  --choices-json artifacts/data-pipeline/sanguo-rag/extracted/resolution-loop/unresolved-triage-choices.json \
  --limit 20 \
  --compact

node tools_node/agent-clis/3klife-sanguo-term-lookup.js \
  --label 孔明 \
  --label 子敬 \
  --label 主公 \
  --json

When labels are not obvious, first generate a research brief:

python server/npc-brain/pipelines/sanguo-rag/generate_term_research_brief.py --top 30

Then delegate the brief to Sanguo Term Researcher if available. Ask it to return answer, confidence, evidence, and a suggested personRecord only for clear people.

If no web-capable agent/tool or reusable CLI is available in the current environment, do not pretend a web check was performed. Provide the research brief path and continue deterministic loop work.

Research Evidence Rules

• 先檢查《三國演義角色列表》：https://zh.wikipedia.org/wiki/三國演義角色列表，可快速確認稱呼是否屬於《演義》人物名單。
• Prefer cross-checking at least two sources for person claims.
• Good sources include 三國志人物列表、三國演義人物列表、維基百科 / 維基文庫 / Wikisource, 中國哲學書電子化計劃, 萌典/漢典 for common terms, and reliable encyclopedia pages.
• Corpus evidence alone proves the label exists in the text; it does not prove whether it is person/place/noise.
• For person, output generalId in kebab pinyin style and a conservative faction (wei, shu, wu, enemy, neutral).
• For geography, offices, collective nouns, book phrases, or segmentation artifacts, choose B noise.
• For uncertain historical/literary ambiguity, choose C ambiguous.

Stop Condition

Stop a loop round when:

• A fresh MCQ file was generated, or
• Filled answers were applied and a new loop was verified, or
• A real blocker occurred and is documented.

Always report the current counts: resolved, unresolved, excluded, reviewPending, and number of generated questions.