P
pwallace/mnemotron-wiki-r
A Claude-powered knowledge base for managing a research corpus.
○checked 2026-11-20 — 命令里的包名与条目不一致,未作为可复制命令展示。
更新 2026-11-16
在 GitHub 查看 ↗链接已于 2026-11-17 核查 · HTTP 200 · 打开的是厂商自己的官网
装完之后怎么算装好了
- 装完该看到什么 — 命令正常结束时退出码为 0,终端里没有 "command not found" 或 "permission denied"。
- 一行自检 — 按该仓库 README 里给出的最小用法跑一次;能按 README 的描述输出就是通了。
- 答不出来怎么办 — 该仓库 README 里有排错 / 常见问题段落 —— 先看那里。
安装命令取自该仓库;以上判据只描述可观察到的结果,不替代项目自己的文档。
- ★ 数取值
- 0 ★ · pwallace/mnemotron-wiki-r · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-17 · 最近见到 2026-11-20 · 共见到 4 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 个来源 · repo
- 安装
- 安装命令见上方「安装与配置」。
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
pwallace/mnemotron-wiki-r;我们收录于 2026-11-17;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。README 原文(英文,未翻译) · 在 GitHub 读完整 README →
# Mnemotron Wiki for Research
以下摘录自该仓库的 README(片段)
> Created with assistance from Claude Code (Claude Sonnet 5.5); may not have been fully reviewed by a human.
> Copyright (C) 2026 Patrick R. Wallace, Hamilton College LITS. Code and Claude instructions: GPL-3.0-or-later. Documentation: GFDL-1.3-or-later. See License.
A Claude-powered knowledge base for managing a research corpus. Drop source
documents into raw_ingest/ and Claude processes them into a structured,
cross-referenced markdown wiki — extracting text from native PDFs and web
pages, and transcribing scanned documents through an offline-first OCR
pipeline. All source content is retained in markdown; raw scan files and
PDFs are discarded after ingest.
See 1.1 Claude backends,
1.2 Concurrency and parallel agents,
1.3 Relationship to the sibling tool and
AUTHENTICATION.md.
## Contents
1. Overview
2. System requirements
3. Installation
4. Day-to-day workflow
5. Supported formats
6. OCR pipeline
7. Wiki structure
8. Configuration reference
9. Script reference
10. Exporting to HTML
11. Troubleshooting
12. License
## 1. Overview
Mnemotron Wiki for Research turns a pile of heterogeneous research documents
into a navigable, cross-referenced knowledge base. The central design
principles are:
Offline-first OCR. Tesseract runs locally for printed text. Claude Vision
is used only when Tesseract's output fails a quality check on a given page, or
when the content is handwritten. A lightweight thumbnail pre-flight avoids
wasting time running full-resolution Tesseract on documents it cannot read.
Retain in markdown, discard originals. After ingest, each document's
content is preserved as a markdown page in wiki/sources/. Raw scan images,
PDFs, and other originals are deleted. The wiki/ folder is the canonical
archive.
Synthesis over filing. Claude does not merely file documents — it reads
each source and updates or creates thematic pages in wiki/topics/ that
synthesize findings across multiple sources, with inline citations back to the
source pages.
Idempotent manifest. A content-hash manifest (.manifest.json) ensures no
file is ever processed twice, even if renamed or moved within raw_ingest/.
Internet Archive integration. ia_ingest.py fetches items directly from
Internet Archive by identifier, preferring IA's pre-built Tesseract OCR text
(_djvu.txt) and falling back to a local OCR pass on the original PDF when
the pre-built text is absent or too noisy.
### 1.1 Claude backends: headless Claude Code or the API
Every script that calls Claude goes through scripts/llm.py, which has two
interchangeable backends:
| Backend | How it works | Credentials |
| cli (default) | Runs the headless Claude Code CLI (claude -p ...) | Whatever login the CLI already has (subscription, Team/Enterprise SSO, or a long-lived token from claude setup-token). No API key or SDK needed. |
| api | Calls the Anthropic Messages API through the anthropic SDK | An API key via ANTHROPIC_API_KEY, MNEMOTRON_API_KEY_FILE, MNEMOTRON_API_KEY_COMMAND, or another credential source the SDK supports. |
Choose with MNEMOTRON_BACKEND=cli|api|auto (or {"llm": {"backend": "..."}} in
an untracked config.local.json). With auto (the default) the CLI is used
whenever claude is on your PATH; the API is used only if you ask for it or
> Copyright (C) 2026 Patrick R. Wallace, Hamilton College LITS. Code and Claude instructions: GPL-3.0-or-later. Documentation: GFDL-1.3-or-later. See License.
A Claude-powered knowledge base for managing a research corpus. Drop source
documents into raw_ingest/ and Claude processes them into a structured,
cross-referenced markdown wiki — extracting text from native PDFs and web
pages, and transcribing scanned documents through an offline-first OCR
pipeline. All source content is retained in markdown; raw scan files and
PDFs are discarded after ingest.
See 1.1 Claude backends,
1.2 Concurrency and parallel agents,
1.3 Relationship to the sibling tool and
AUTHENTICATION.md.
## Contents
1. Overview
2. System requirements
3. Installation
4. Day-to-day workflow
5. Supported formats
6. OCR pipeline
7. Wiki structure
8. Configuration reference
9. Script reference
10. Exporting to HTML
11. Troubleshooting
12. License
## 1. Overview
Mnemotron Wiki for Research turns a pile of heterogeneous research documents
into a navigable, cross-referenced knowledge base. The central design
principles are:
Offline-first OCR. Tesseract runs locally for printed text. Claude Vision
is used only when Tesseract's output fails a quality check on a given page, or
when the content is handwritten. A lightweight thumbnail pre-flight avoids
wasting time running full-resolution Tesseract on documents it cannot read.
Retain in markdown, discard originals. After ingest, each document's
content is preserved as a markdown page in wiki/sources/. Raw scan images,
PDFs, and other originals are deleted. The wiki/ folder is the canonical
archive.
Synthesis over filing. Claude does not merely file documents — it reads
each source and updates or creates thematic pages in wiki/topics/ that
synthesize findings across multiple sources, with inline citations back to the
source pages.
Idempotent manifest. A content-hash manifest (.manifest.json) ensures no
file is ever processed twice, even if renamed or moved within raw_ingest/.
Internet Archive integration. ia_ingest.py fetches items directly from
Internet Archive by identifier, preferring IA's pre-built Tesseract OCR text
(_djvu.txt) and falling back to a local OCR pass on the original PDF when
the pre-built text is absent or too noisy.
### 1.1 Claude backends: headless Claude Code or the API
Every script that calls Claude goes through scripts/llm.py, which has two
interchangeable backends:
| Backend | How it works | Credentials |
| cli (default) | Runs the headless Claude Code CLI (claude -p ...) | Whatever login the CLI already has (subscription, Team/Enterprise SSO, or a long-lived token from claude setup-token). No API key or SDK needed. |
| api | Calls the Anthropic Messages API through the anthropic SDK | An API key via ANTHROPIC_API_KEY, MNEMOTRON_API_KEY_FILE, MNEMOTRON_API_KEY_COMMAND, or another credential source the SDK supports. |
Choose with MNEMOTRON_BACKEND=cli|api|auto (or {"llm": {"backend": "..."}} in
an untracked config.local.json). With auto (the default) the CLI is used
whenever claude is on your PATH; the API is used only if you ask for it or
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/pwallace-mnemotron-wiki-r/)
- 来源
- github · https://github.com/pwallace/mnemotron-wiki-r
- 按领域浏览
- 开源项目
- 收录于
- 今天
更多条目
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。