Forge raw sources into an agent-ready knowledge OS.
把原始素材锻造成 AI Agent 可长期使用的知识操作系统。
WikiForge 是一个面向 AI Agent 的本地知识库模板。它把原始资料、外部参考、可信事实、Agent 规则和自检脚本分开，让 AI 可以长期读取、维护、审计同一个知识系统。
WikiForge is a local knowledge-system template for AI agents. It separates raw sources, external references, trusted facts, agent rules, and validation scripts so agents can use the wiki over time.
很多 AI 知识库的问题不是“资料太少”，而是资料进入系统后很快混在一起：原始笔记像事实，外部观点像结论，AI 第二次回答时已经分不清哪些是验证过的。
Many AI knowledge workflows fail because raw notes, external claims, and verified decisions collapse into the same layer.
WikiForge 用几个简单边界解决这个问题：
WikiForge keeps the layers separate:
sources/ raw captured material
domains/ distilled working knowledge and external references
wiki/ index, facts, log, evaluation set, and agent-facing checks
AGENTS.md operating agreement for AI agents
tools/ local lint, recall, exam, ingest, and redaction scripts
先跑脱敏检查，再跑 wiki 检查，然后试一次本地 ingest、reference recall 和 knowledge exam。
Run the redaction gate first, then lint, ingest a sample, recall references, and run the regression exam.
python tools\redaction_check.py
python tools\wiki_lint.py
python tools\ingest_stub.py examples\sample-source.md
python tools\wiki_radar.py query "customer interviews pricing"
python tools\knowledge_exam.py regression完整步骤见 docs/quickstart.md。
Full walkthrough: docs/quickstart.md.
-
原始材料放在
sources/，不要直接改成结论。 -
外部观点先放在
domains/<area>/references/，不要直接升级成事实。 -
wiki/facts.md只接受verified和hypothesis。 -
hearsay可以归档和召回，但不能进入事实表。 -
tools/redaction_check.py是发布阻断项；失败时不要 publish、push 或 package。 -
Raw material stays in
sources/. -
External claims stay in
domains/<area>/references/until validated. -
wiki/facts.mdonly acceptsverifiedorhypothesis. -
hearsaymay be archived and recalled, but must not become operating truth. -
tools/redaction_check.pyis a release gate.
这个仓库只包含可复用模板、虚构示例和轻量本地脚本。它不包含真实业务数据、message-platform 凭证、bot 名称、session 文件、客户信息或生产自动化。
This repository contains only reusable templates, fictional examples, and lightweight local scripts. It does not include private data, credentials, bot names, sessions, customer data, or production automation.