PDF Brain — Research → Practical System Moves

Use this skill when the user wants evidence-backed synthesis from the docs library (600+ books, PDFs, long-form references), not generic web summarization.

Pipeline v2 (ADR-0234)

The docs pipeline uses a staged artifact chain:

• Extraction: opendataloader-pdf → structured markdown with headings, tables, reading order
• Chunking: markdown-native heading detection, no overlap, hierarchical section + snippet chunks
• Embeddings: nomic-embed-text via ollama GPU (768-dim, retrieval-tuned, pre-computed at ingest) in docs_chunks_v2 collection
• Artifacts: durable on NAS at /Volumes/three-body/docs-artifacts/{docId}/ — .md, .meta.json, .chunks.jsonl
• Summaries: LLM-generated per-document summaries in .meta.json

When to Use

Trigger cues (explicit or implied):

• "research this" / "from the library" / "from the books"
• "pdf brain" / "correlate this to our system"
• "what does the research say" / "what do the books say"
• "expand this into practical ideas"

Retrieval Workflow

CLI path (preferred for interactive sessions)

# Search across all books — semantic by default (nomic 768-dim)
joelclaw docs search "distributed consensus" --limit 8

# Search within a specific book
joelclaw docs search "consensus" --doc designing-dataintensive-applications-39cc0d1842a5

# Expand a chunk into surrounding context
joelclaw docs context <chunk-id> --mode snippet-window --before 2 --after 2

# Get the full parent section
joelclaw docs context <chunk-id> --mode parent-section

# Get neighboring sections for broad context
joelclaw docs context <chunk-id> --mode section-neighborhood --neighbors 2

# Read the full structured markdown of a book
joelclaw docs markdown <doc-id>

# Get document summary + taxonomy metadata
joelclaw docs summary <doc-id>

API path (for programmatic access or docs-api consumers)

GET /search?q=distributed+consensus&semantic=true&expand=true&assemble=true
GET /docs/:docId/toc
GET /docs/:docId/markdown
GET /docs/:docId/summary
GET /chunks/:chunkId

The docs-api runs on k8s at docs-api:3838 (Bearer auth required).

Context expansion strategy

The library supports progressive context expansion:

1. Search → chunk-level hits with heading_path and snippet
2. snippet-window → 2 chunks before/after for local context
3. parent-section → the full section containing the snippet
4. section-neighborhood → adjacent sections for broader flow
5. markdown → the complete structured book text

Start narrow, expand only when needed. Don't dump full books into context.

Evidence Synthesis

Build an evidence ledger

While reading, keep a compact ledger:

• doc (title)
• chunk-id
• claim (one sentence)
• relevance (why it matters to this problem)

Never output synthesis without traceable evidence.

Convert evidence into principles

Turn each claim into an operational principle in imperative form:

• "Treat partial failure as normal."
• "Fail fast at dependency boundaries."
• "Prefer idempotent replay-safe remediation loops."

Avoid vague advice. Each principle must imply a technical behavior.

Correlate to joelclaw philosophy

Map principles to existing joelclaw operating rules:

• single source of truth
• silent failures are bugs
• Inngest durability + retries
• CLI-first agent interface
• observability required at every step
• skill/doc updates when reality changes

Translate into action

For each principle, produce:

1. Concrete change (file/service/config path)
2. Validation gate (exact command)
3. Failure signal (what proves it did not work)
4. Rollback or containment move

Taxonomy

The library is classified via SKOS taxonomy:

• jc:docs:programming (systems, languages, architecture)
• jc:docs:business (creator economy)
• jc:docs:education (learning science, pedagogy)
• jc:docs:design (game, systems, product)
• jc:docs:marketing, jc:docs:strategy, jc:docs:ai, jc:docs:operations

Use --concept jc:docs:programming:systems to narrow by domain. Use joelclaw docs status to see facet counts per concept.

Rules

• Do not fabricate quotes or claims.
• Always cite chunk IDs for non-obvious assertions.
• Do not output "book report" fluff. Translate to operations.
• If infra changes are proposed, include verification commands.
• If work implies architectural policy change, tie it to an ADR path.
• Start with search, expand only as needed. Don't waste context on full book dumps.