reference-import

Reference Import Skill

Import git repos, web crawls, documents, books, and YouTube videos as topic-scoped references into the knowledge base. All sources go through a two-step flow: convert (produce staging dir with markdown) then import (index into the knowledge base).

Decision Flow

Follow this order — stop as soon as you have an answer:

1. Search first: knowledge_search(query="<topic>", categories=["references"]). If results exist, use them to answer. Done.
2. Git import (preferred for whole doc sets): find the docs repo via gh, convert with ghost convert git, then import. Use background: true for the convert step.
3. Crawl import (fallback): only if no git source exists (e.g. docs-only site).
4. YouTube import: if the OPERATOR wants a single video, convert with ghost convert youtube, then import the staging dir and run the video-import agent.
5. PDF/Document import: download first (curl), then ghost convert pdf, then import.
6. After starting the background convert: tell the OPERATOR it's importing, include any other pending responses (project offers, plans, etc.), then end your turn. A follow-up turn is triggered automatically when the convert finishes — you'll see the [shell-command completed] system message. Read the staging dir, pick a topic, then run the import. Note: reference records appear in the DB almost immediately; only the embeddings trail behind.

Two-Step Flow

Step 1: Convert

Convert produces a staging directory with markdown files and prints the staging path to stdout:

ghost convert git <url> [--paths dir1,dir2] [--extensions .md,.rs] [--git-ref <ref>]
ghost convert crawl <url> [--max-depth 3] [--max-pages 50]
ghost convert youtube --url https://www.youtube.com/watch?v=<id>
ghost convert pdf <path> [--no-ocr] [--page-range "1-10"] [--timeout 900]

The staging directory defaults to <workspace>/.staging/<slug>/. The command prints the staging path and provenance details (source URL, git ref, etc.) to stdout — capture these for the import step.

YouTube import is v1 single-video only. Transcript acquisition priority is: manual captions, then auto captions, then CPU Whisper fallback if captions are missing. The converter splits transcripts into multiple markdown files based primarily on length, not uploader chapters, so the reference files stay readable for rereading and search.

Step 2: Inspect

After the convert completes, read a few files from the staging dir to understand the content, then pick an appropriate --topic.

Step 3: Import

ghost reference import <staging-dir> --topic <topic> \
    [--source-type git|crawl|file|youtube] \
    [--source-url <url>] \
    [--version-ref <commit-hash>] \
    [--git-ref <branch-or-tag>]

The import indexes the staging directory into the knowledge base and writes files to references/{topic}/. Pass the provenance flags from the convert output so the import record is fully traceable.

For YouTube imports, pass --source-type youtube and the original video URL from the converter output. The import only needs the markdown transcript sections in staging; there is no need to keep audio or subtitle artifacts.

YouTube Import

1. Convert (single-video URLs only in v1):

ghost convert youtube --url https://www.youtube.com/watch?v=<id>

The converter tries manual captions first, auto captions second, and CPU Whisper only when captions are unavailable. It prints the staging path plus metadata such as the source URL, video ID, title, channel, transcript source, section count, and chapter count.

2. Import:

ghost reference import /path/to/.staging/<slug> --topic videos/<slug> \
    --source-type youtube \
    --source-url https://www.youtube.com/watch?v=<id>

3. Create notes — use the agent tool to start the video-import agent. Pass the topic, title, and channel from the convert output as args:

{
  "action": "start",
  "agent": "video-import",
  "prompt": "Create notes from the imported video.",
  "args": {
    "topic": "videos/<slug>",
    "title": "<title from convert output>",
    "channel": "<channel from convert output>"
  }
}

Use video-import for transcript-based notes. It is tuned for shorter section files, section-aware reading, and timestamped source notes.

Git Import (Preferred)

Finding the docs repo

Documentation often lives in a separate repo (e.g. DioxusLabs/docsite, not DioxusLabs/dioxus). One search is enough:

gh search repos "docs OR docsite OR website" --owner=<Org> --json name,description

Choosing paths + extensions

Browse the repo to pick the narrowest --paths:

gh api repos/<owner>/<repo>/contents/ --jq '.[].name'

For docs repos use --extensions .md; for code examples add .rs, .py, etc. Omit --paths only for small repos.

Running the convert (background)

Git converts clone the repo and embed every file, which is slow on CPU. Always use background mode (background: true) for the convert step:

ghost convert git https://github.com/DioxusLabs/docsite --paths docs-src/0.7/src --extensions .md

Tell the OPERATOR: "I'm converting the Dioxus docs in the background — I'll import them once conversion finishes." Finish any other pending responses, then end your turn. The completion watcher triggers a follow-up turn automatically.

After convert completes

When you see [shell-command completed], the stdout contains the staging path. Read a few files to pick a good topic name, then run the import:

ghost reference import /path/to/.staging/docsite --topic dioxus/docs \
    --source-type git --source-url https://github.com/DioxusLabs/docsite --git-ref main

Crawl Import (Fallback)

Use only when no git source exists. Run with background: true:

ghost convert crawl https://docs.example.com/ --max-depth 2 --max-pages 30

After convert completes, import the staging dir:

ghost reference import /path/to/.staging/docs-example-com --topic example/docs \
    --source-type crawl --source-url https://docs.example.com/

PDF / Document Import

Why download first?

Many sites serve consent pages, CAPTCHAs, or redirects instead of the actual file. Downloading first lets you verify the file is real (check size, file type) before converting. It also keeps the original in uploads/ for re-import if needed.

Workflow

1. Download (foreground — verify the file is real before converting):

curl -L -o uploads/rulebook.pdf 'https://example.com/rulebook.pdf'

Check file size and type — some sites return HTML error pages with HTTP 200.

2. Convert (background — OCR can take minutes, use background: true):

ghost convert pdf uploads/rulebook.pdf

3. Import after convert completes:

ghost reference import /path/to/.staging/rulebook --topic boardgames/arknova \
    --source-type file

PDF convert options (use ONLY when explicitly requested)

These are optimization overrides. Use defaults unless the OPERATOR asks otherwise.

Flag	Default	When to use
--no-ocr	OCR on	OPERATOR says PDF is digital, wants speed
--page-range "1-10"	full doc	OPERATOR wants specific pages only
--timeout 900	config	OPERATOR needs more time for huge docs

Do NOT guess at these options. Do NOT add --no-ocr to "speed things up". The OPERATOR will tell you if they want non-default behavior.

Uploaded files

When the OPERATOR uploads a file, it lands in uploads/ in the workspace. Convert it directly — no curl needed:

ghost convert pdf uploads/<filename>

After import, the original file is preserved in references/<topic>/_originals/. Clean up the uploaded file:

rm uploads/<filename>

Book Import (EPUB)

Workflow

1. Convert (fast — pure Rust, no external deps):

ghost convert epub --path /path/to/book.epub

The output shows the staging path, title, authors, and chapter count.

2. Import:

ghost reference import /path/to/.staging/book-slug --topic books/<slug> \
    --source-type book

Pick a descriptive --topic under books/ (e.g., books/fiction, books/economy).

3. Create notes — use the agent tool to start the book-import agent. Pass the topic, title, and authors from the convert output as args:

{
  "action": "start",
  "agent": "book-import",
  "prompt": "Create notes from the imported book.",
  "args": {
    "topic": "books/<slug>",
    "title": "<title from convert output>",
    "authors": "<authors from convert output>"
  }
}

Files on Disk

All imported references are written to disk under references/{topic}/ in the workspace AND stored in the DB for search. This means you can file_read on paths returned by knowledge_search to get full reference content — no need to re-fetch from the web.

• Git imports: files mirror the repo structure (references/{topic}/{rel_path})
• Crawl imports: filenames are URL slugs (references/{topic}/{slug}.md)
• PDF imports: one markdown file per source document
• YouTube imports: one markdown file per transcript section
• knowledge_search results for references include a path: field you can file_read

Staging Directory

The staging directory (.staging/ in workspace) is auto-cleaned after a successful import. If you need to re-import with different options, re-run the convert step.

Post-Import: Enrich the Topic Note

If the topic is new, after import, a placeholder note exists at notes/<topic>/index.md. Edit it with a meaningful description — what the library/document covers, key concepts. This makes the topic discoverable via semantic search.

Topic Hierarchy

Topics are namespaces: dioxus, dioxus/docs, dioxus/source.

• Searching with topic="dioxus" finds results across all sub-topics
• Each import writes _import.toml alongside the references

Post-Import Search

knowledge_search(query="hooks", topic="dioxus", categories=["references"])

Cleanup

ghost reference delete --topic dioxus/docs

Updating References

When the OPERATOR asks to refresh or update existing reference material, or when you notice imported docs may be stale (e.g. a library released a new version):

ghost reference update --topic <name> [--ref <tag-or-branch>]

This re-fetches from the original source and applies changes:

• New files are added
• Changed files are updated (content + embeddings)
• Files deleted upstream are removed — unless cited by notes, in which case they're moved to _orphaned/ with a warning

For git sources, the command short-circuits if the upstream commit hash hasn't changed. Use --ref to switch to a different branch or tag.

Examples (use background: true):

ghost reference update --topic dioxus/docs
ghost reference update --topic dioxus/docs --ref v0.6

CLI Reference

# Convert
ghost convert git <url> [--paths <paths>] [--extensions <exts>] [--git-ref <ref>] [--output <dir>]
ghost convert crawl <url> [--max-depth 3] [--max-pages 50] [--output <dir>]
ghost convert pdf <path> [--no-ocr] [--page-range "1-10"] [--timeout <secs>] [--output <dir>]
ghost convert epub --path <path> [--output <dir>]

# Import
ghost reference import <staging-dir> --topic <name> \
    [--source-type git|crawl|file] \
    [--source-url <url>] \
    [--version-ref <commit-hash>] \
    [--git-ref <branch-or-tag>]

# Manage
ghost reference update --topic <name> [--ref <tag-or-branch>]
ghost reference delete --topic <name>
ghost topics list