crimeacs/claude-note/ingest-2

Document ingestion

Use the installed claude-note ingest command to preserve the vault's writing, provenance and merge behavior. The CLI accepts one local .pdf, .docx, .md, or .txt file per invocation. It calls the local Claude CLI for extraction.

claude-note ingest "/path/to/paper.pdf"
claude-note ingest "/path/to/runbook.md" --internal --title "Service Runbook"
claude-note ingest "/path/to/paper.pdf" --dry-run

Resolve the configured vault before writing. CLAUDE_NOTE_VAULT_ROOT overrides vault_root in ${XDG_CONFIG_HOME:-~/.config}/claude-note/config.toml. claude-note health reports configuration; claude-note ingest --help lists supported flags. If no destination is configured or inferable from the user's request, ask for the destination. Routine ingestion and evidence-backed merges within the authorized vault do not require another permission prompt.

Sources and output

For a URL, fetch its readable content, save an explicitly named local document, and include the original URL in that source. Then pass the saved file to the CLI. Retain the URL on the resulting source note. Do not pass a URL directly to the CLI or claim the CLI fetches web pages. For several files, invoke the command once per file.

Literature mode writes literature/lit-*.md with type: literature; internal mode writes internal/int-*.md with type: source. Each document has a source index and source-linked concept notes. Use the amount of knowledge the document supports; no fixed concept count is a quality target. Keep source findings, interpretation and unanswered questions distinguishable. Add only verified existing relationships, preserving project boundaries and human annotations.

Read literature-format.md for external sources or internal-format.md for internal material. Sample output illustrates the source/concept pair.

Existing knowledge

Exact filenames are checked first. Semantic ingestion candidates require QMD enabled = true, an explicit collection for this vault, and ingest_dedup_enabled = true in [qmd]. This optional feature uses vector retrieval and may load an embedding model. A search rank is a candidate signal, not proof of equivalence: the CLI resolves the result to a real file in the configured output directory and asks Claude whether source-backed information is worth adding. An unrelated collection or a source index is never a concept merge target. When QMD is disabled or unavailable, exact filename handling and new-note creation still work.

The source index links to actual created or merged concepts. Existing source indexes update a managed block; other human text is preserved. If a filename is already owned by another source or an unrelated note, report the conflict and choose an appropriate distinct source title within the user's scope. Never replace human content merely because titles or scores are similar.

Finish

Inspect the returned paths and source backlinks. Summarize what was created, merged or skipped, and any extraction/merge failures. A dry run extracts and previews without writing vault notes; it still invokes the extraction model. Do not claim semantic candidates were searched during a dry run: merging happens only during the write phase. Publication or external sharing is a separate act.