AutoWiki Ingest

The EN→PI→R loop for document ingestion. Process one document at a time. For batch runs (5+ docs), run a watchdog session in parallel.

Before you start

• Wiki must exist with _navigate.md, SCHEMA.md, index.md, log.md.
• Read _navigate.md to understand directory layout and conventions.
• Read index.md for existing pages and log.md (last 30 lines) for recent activity.

State tracking

Use $WIKI/state.json to survive context resets:

{"current_doc": "...", "chunk": 0, "total_chunks": N,
 "extracted": [], "pages_touched": [], "stale_count": 0, "last_seen": <ts>}

Read state at session start. Update after each chunk. If state exists, resume from where it left off.

Step 1: Capture raw source

First, check if already ingested. Compute sha256 of the document content. Search existing raw/ files for this sha256:

grep -rl "sha256: <digest>" "$WIKI"/raw/ 2>/dev/null

If found and the user didn't explicitly request re-ingestion, skip this document. Report: "Already ingested as <path>. Skipping."

If no match, proceed:

1. Read the document. Use read_file for text/code/logs. For PDFs,

extract text first: python3 -c "import fitz; ..." or paste content.

1. Save to raw/<subdir>/<descriptive-slug>.md with frontmatter:

---
source_url: <url or file path>
ingested: YYYY-MM-DD
sha256: <hex>
type: article|paper|code|log|data
---
<content>

Compute sha256: python3 -c "import hashlib,sys; print(hashlib.sha256(sys.stdin.read().strip().encode()).hexdigest())" < <file>

Subdir: PDFs → papers/, code → code/<repo>/, logs/CSV → logs/, rest → articles/.

Step 2: Chunk with strategy

Load autowiki-chunk skill for chunking strategy and instructions. Pick strategy by file extension:

Load ONE chunk at a time. Process it fully before loading the next. To read specific lines: read_file <file> offset=<N> limit=<M>.

Implementation-only code blocks that yield no extractable concepts are normal — don't count them as stale. Only increment stale_count when a chunk yields ZERO items.

Step 3: EN — Extract (per chunk)

From the current chunk, pull out:

• Entities: named things (people, orgs, systems, tools, repos)
• Concepts: ideas, techniques, patterns, findings
• Claims: assertions of fact with numbers or comparisons
• Relationships: entity A relates to entity B via concept C

Note the provenance: which file, which section/line numbers.

Write extracted items to $WIKI/.tmp/extracted.json (append mode, one JSON object per line).

If a chunk yields ZERO extractable items:

• Mark it: increment stale_count in state.json
• stale=1: retry same chunk, same strategy
• stale=2-3: halve the chunk size and retry. Don't dig deeper.
• stale≥4: flag this document for human review, skip to next doc

Important: implementation-only code blocks that yield no extractable concepts are normal for the code strategy. Only increment stale_count when a block yields ZERO items. Processing details without extracting new knowledge is expected, not a stall.

If a block references entities from earlier blocks (e.g., function using FEATURE_NAMES from imports block), search the current document's already-extracted items before creating duplicate pages. Two functions doing the same thing = one concept page. Update the existing page rather than creating a new one.

If extraction succeeds: proceed to PI phase. Don't touch stale_count. Update state.json: increment chunk counter, set last_seen timestamp.

Step 4: PI — Understand and integrate

For each extracted item (entity, concept, claim):

1. Search KB for existing page:

   search_files "<name>" path=$WIKI file_glob="*.md" output_mode="files_only"

1. If matching page exists: read it, update with new information,

add the new source to frontmatter sources: list, add cross-references. Bump updated date. Write back.

1. If no match: create new page using the template below.

File at entities/<slug>.md or concepts/<slug>.md or claims/<slug>.md.

1. Every page must have YAML frontmatter:

   ---
   title: Page Title
   created: YYYY-MM-DD
   updated: YYYY-MM-DD
   type: entity|concept|claim
   tags: [from SCHEMA.md taxonomy]
   sources: [raw/path/to/source.md]
   confidence: high|medium|low
   contested: false
   contradictions: []
   ---

1. Minimum 2 `[[wikilinks]]` per page. Link to related entities, concepts.

Pages without links are invisible in the knowledge graph.

1. Track touched pages: add slug to pages_touched in state.json.

Step 5: R — Validate (per chunk)

Before moving to next chunk:

1. Source fidelity: for each claim, re-read the raw source section.

Compare numbers, qualifiers, context. If claim text doesn't match source, set confidence: low and add a note.

1. Internal consistency: search claims/ for conflicting claims:

   search_files "<topic>" path=$WIKI/claims file_glob="*.md" output_mode="files_only"

If found: mark both contested: true, add contradictions: [other-slug], set confidence: low on the newer claim. Do NOT auto-resolve.

1. Confidence scoring:

• high: 2+ independent sources, no contradictions, fidelity verified
• medium: single source, no contradictions (default for new claims)
• low: contradiction exists, fidelity issue, or single-source opinion

Step 6: After all chunks

1. Load autowiki-index skill to rebuild index and log.
2. Run independent audit: pick 3 claims at random. Re-read their raw sources.

If audit disagrees with original confidence, mark contested: true.

1. Run drift check on all raw sources:

   for f in "$WIKI"/raw/**/*.md; do
     stored=$(grep '^sha256:' "$f" | cut -d' ' -f2)
     body=$(sed '1,/^---$/d' "$f" | tail -n +2)
     current=$(echo "$body" | python3 -c "import hashlib,sys; print(hashlib.sha256(sys.stdin.read().strip().encode()).hexdigest())")
     [ "$stored" != "$current" ] && echo "DRIFT: $f"
   done

1. Clean up $WIKI/.tmp/ and $WIKI/state.json.
2. Report: pages created, updated, contradictions found, drift detected.

Watchdog (run in parallel for batch jobs)

In a separate session, check every hour:

state=$(cat "$WIKI/state.json" 2>/dev/null)

• No state file → run hasn't started.
• last_seen > 2h ago → run is dead. Restart from chunk in state.
• stale_count ≥ 2 → run is stuck. Nudge by injecting alternative chunk size.
• stale_count ≥ 4 → flag for human.

The watchdog never reads document content or wiki pages. Only state.json.

Never

• Load more than one KB page into context at once.
• Load the full index.md — scan search_files output.
• Auto-resolve contradictions.
• Modify raw/ files.
• Let the watchdog read task data.