AI Research Wiki

An LLM-maintained knowledge base for AI research, following the LLM Wiki pattern. Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.

Browse the wiki

Structure

sources/          Raw source documents synced from Readwise Reader (gitignored)
summaries/        LLM-generated summaries of each source (committed)
                  Contains original URL, summary, and main ideas
wiki/             LLM-generated knowledge pages (summaries, entities, concepts,
                  comparisons, syntheses). LLM owns this entirely.
  index.md        Content catalog of every page, grouped by category
  log.md          Reverse-chronological activity log
CLAUDE.md         Schema defining structure, conventions, and workflows
sync-readwise.py  Pulls AI-tagged content from Readwise Reader into sources/
ingest.py         Full pipeline: sync → summarize → ingest → PR
.env              API tokens (not committed)

Setup

1. Add your Readwise token to .env:

   READWISE_TOKEN=rw_xxxxxxxxxxxxx
   

   • Readwise token: https://readwise.io/access_token
2. Ensure claude (Claude Code CLI) is on your PATH — summarization and wiki ingest both run through it. A Claude Code subscription is required.

Running the Pipeline

python3 ingest.py              # sync + summarize + ingest + open PR
python3 ingest.py --skip-sync  # skip Readwise sync (sources already up to date)

ingest.py runs the full pipeline:

1. Sync — pulls AI-tagged documents from Readwise Reader into sources/ (incremental, stdlib only, tracks state in sources/.sync-state.json)
2. Summarize — generates summary docs in summaries/ via the Claude CLI (claude-haiku-4-5-20251001), parallelized up to 5 concurrent requests; skips sources whose content hasn't changed
3. Ingest — runs Claude Code to read summaries and create/update wiki pages, cross-references, index, and log
4. PR — commits changes and opens a pull request

Early exit (no branch created) if nothing is new after sync.

Recovery

If a run fails partway through — after summarize but before ingest or PR — you can resume from the ingest step without re-syncing or re-summarizing:

git checkout <the ingest branch>   # branch created during the failed run
python3 ingest.py --resume

--resume skips steps 1–2, runs the wiki ingest against the existing summaries, then commits, pushes, and creates or updates the PR. Use it when you see summaries in summaries/ but no (or an incomplete) PR.

Using the Wiki

Open this folder as an Obsidian vault. All pages use standard markdown links for cross-references, which work on both GitHub and Obsidian. The graph view shows how everything connects.

Operations

All operations are performed by asking the LLM (Claude) in conversation:

• Ingest — Run python3 ingest.py for the full automated pipeline, or ask Claude directly to ingest summaries. Claude reads the summaries, creates/updates wiki pages, and maintains cross-references, the index, and the log.
• Query — Ask questions about the content. Claude synthesizes answers from wiki pages with citations. Good answers can be filed as new wiki pages.
• Lint — Ask Claude to audit the wiki for contradictions, stale claims, orphaned pages, or gaps.

Conventions

• File names: lowercase, hyphen-separated (e.g., transformer-architecture.md)
• Each page starts with an H1 title
• Standard markdown links (work on both GitHub and Obsidian):
  • Wiki-to-wiki: [Page Title](page-name.md)
  • Wiki-to-source: [Source Title](../summaries/filename.md)
• YAML frontmatter tags where helpful (e.g., tags: [llm, architecture])
• Log entries are reverse-chronological (newest first)