Wiki Ingestion

Hermes Wiki Ingestion

This is the default ingestion skill for Hermes LLM Wikis. It describes how raw material becomes attributable Wiki Pages. A wiki can override this default with hermes wiki skills set ingestion <skill-name>.

Media sources (PDFs, images, audio, video, YouTube, social links) have their own protocols — load wiki:wiki-media-ingestion before ingesting them.

The pipeline

Every ingest follows the same deterministic flow:

1. Snapshot — the source bytes are stored append-only under raw/ (never overwritten; re-ingesting changed content creates a new versioned snapshot and flags drift_detected).
2. Classify — built-in classifiers detect article, paper, or transcript. If none match, trusted custom classifiers (see wiki plugins) get a chance. Anything else lands as unknown and is retained for review, never silently dropped.
3. Process — the processor for the classified label generates or updates Wiki Pages (a source page plus extracted concept/entity pages), cross-linking existing pages.
4. Propagate — index.md sections, log.md attribution rows, the SQLite projection, and a git commit are all updated atomically; failures roll back.

Ingesting one source

hermes wiki ingest <local-path-or-https-url> [--wiki <slug>] [--author <name>]

• Requires a write grant (HERMES_WIKI=<slug> or config write grant).
• URLs are fetched and deduplicated by sha256 — unchanged content is a no-op (no change: <source-id>).
• The 50MB Phase-1 cap applies; larger sources are refused.

The inbox flow

Drop files into <wiki-root>/raw/inbox/ (or let monitors deliver them), then:

hermes wiki inbox                  # list pending files with status + classifier
hermes wiki ingest --inbox         # batch-process everything processable

Per-file outcomes:

• Ingested <name> class=<label> — pages were generated.
• Retained <name> class=unknown — kept in the inbox for a human/agent to classify. Override via the dashboard inbox view or the classify API, then re-process.
• Skipped <name> status=oversized — exceeds the 50MB cap; not processable in this phase. Delete it or handle it out-of-band.

Classifier overrides persist in raw/inbox_status.json and survive failed attempts; an override pins the label for the next ingest --inbox run.

Recurring ingestion (monitors)

For sources that update on a schedule, define a portable monitor instead of ingesting manually:

hermes wiki monitor --source rss --schedule "0 8 * * *" --prompt "Sweep feeds…"
hermes wiki monitor --setup --yes

Monitor sweeps run with author_kind=cron and full attribution.

Custom classification

When the built-ins misclassify a domain-specific format, write a classifier at plugins/classifiers/<name>.py exposing classify(source_path) -> ClassLabel | str | None, then trust it explicitly:

hermes wiki plugins trust classifier <name>

Trust is recorded canonically in SCHEMA.md (path + sha256) — editing the file after trusting silently disables it until re-trusted.

Synthesis guidance

When expanding ingestion-derived pages (or writing new ones from ingested sources), apply the synthesis protocol from the wiki-writing skill:

• Page-creation threshold: derive pages for entities/concepts that appear in 2+ sources or are central to this one — never for passing mentions.
• Drift means review: when re-ingest reports drift_detected, review pages whose sources: cite the superseded snapshot. Newer content generally supersedes; genuine conflicts follow the contradiction protocol (contested: true + contradictions: frontmatter). Raw snapshots stay append-only — only wiki pages are updated.

Rules

• Never write into raw/ by hand during ingestion — let the pipeline snapshot.
• Never re-classify by renaming files; use the override surfaces.
• Unknown is a valid outcome: retain, review, re-classify. Do not force-fit.
• Check hermes wiki lint after large batches; it repairs the projection.