Librarian - Semantic Research Skill

Version: 2.0.0 (Protocol-driven)
Status: 🚧 Development
Architecture: Sandwich (🎤 Skill → 👷 Wrapper → ⚙️ Python)

---

What This Skill Does

Search your book library using natural language. Ask questions like "What does Graeber say about debt?" and get precise citations with page numbers.

---

Protocol Flow

CODEBLOCK0

Status: ✅ All nodes ready (v0.15.0 complete)

Protocol Nodes:

1. 1. Load Metadata: Reads .library-index.json + .topic-index.json files
2. Infer Scope: Confidence >75% → proceed | <75% → ask clarification
3. Build Command: INLINECODE2
4. Format Output: Synthesized answer + emoji citations + sources
5. 🤚 Hard Stop: Honest failure > invented answer (VISION.md principle)

Sandwich Architecture:

Flow: 🎤 Skill → 👷 Sh → ⚙️ Py → 👷 Sh → 🎤 Skill

Why this pattern:

1. 1. 🎤 Skill interprets user intent (conversational, flexible, handles ambiguity)
2. 👷 Sh builds correct command syntax (skill errs often, sh hardens protocol)
3. ⚙️ Py executes deterministic work (search, embeddings, JSON output)
4. 👷 Sh formats py output to structured syntax (protocol compliance)
5. 🎤 Skill presents to human (natural language, citations, formatting)

Symbols:

• - 🎤 = Skill (you, AI conversational layer)
• 👷 = Wrapper (librarian.sh, protocol enforcement)
• ⚙️ = Python (research.py, heavy lifting)
• 🤚 = Hard stop (honest failure > invented answer)

---

🤚 Hard Stop Protocol (CRITICAL)

You are a messenger, not the system.

When wrapper returns error codes:

• - ERROR_NO_METADATA → "Não tem metadata. Roda librarian index."
• INLINECODE5 → "Não entendi. Reformula? (topic ou book?)"
• INLINECODE6 → "Sistema quebrado."
• INLINECODE7 → "Não achei nada sobre [query]."

STOP THERE. Do NOT:

• - ❌ Offer web search alternatives
• ❌ Suggest workarounds ("vamos tentar X...")
• ❌ Hallucinate ("maybe the book says...")
• ❌ Apologize or frame as your failure

Hard stop = SUCCESS. You detected system state and reported honestly.

You didn't create the problem. You're just telling the truth:

• - "Tem goteira." ← Bad news, but not your fault.
• "Não tem resultados." ← Reality, not failure.

Reporting hard stops IS your job done. ✅

---

Metadata Structure (Subway Map)

How metadata is organized:

CODEBLOCK1

Navigation:

• - Topic scope = 1 step (scan .library-index.json only)
• Book scope = 2 steps (.library-index.json → infer topics → scan .topic-index.json files)

🔴 CRITICAL: Extension Handling

User NEVER mentions file extensions.

Examples:

• - ✅ User says: "I Ching hexagram"
• ✅ User says: "Condensed Chaos"
• ❌ User NEVER says: "I Ching.epub"

Why: Extension = metadata detail (epub vs pdf), irrelevant to user.

Your job:

1. 1. Match query → book title (NO extension)
2. Pass filename to wrapper (WITH extension: "I Ching.epub")
3. Results show title only (NO extension in output)

Metadata fields:

• - .library-index.json → topics list (big picture)
• INLINECODE14 → books list per topic (narrow view)
• Book metadata: title (user-facing, no ext) + filename (internal, with ext)

Full taxonomy: See backstage/epic-notes/metadata-taxonomy.md

---

How To Use This Skill

Trigger Detection

Activate when user query matches ANY of these patterns:

Book/Author references:

• - "What does [AUTHOR] say about [TOPIC]?"
• "Search [BOOK] for [QUERY]"
• "Find references to [CONCEPT] in [BOOK]"

Topic keywords (with confidence >75%):

• - "tarot", "I Ching", "divination" → chaos-magick
• "debt", "finance", "money", "banking" → finance
• "anarchism", "mutual aid", "commons" → anarchy

Explicit commands:

• - "pesquisa [QUERY]" / "search [QUERY]"
• "procura [CONCEPT]" / "find [CONCEPT]"
• "librarian: [QUERY]"

If confidence <75% → CLARIFY (ask user)

---

Node 2: 🎤 Infer Scope

Determine WHAT to search (topic or book) from user intent.

AI = router. Intelligence is in the index (embeddings). You just match query → scope.

Confidence Logic (Binary)

Read metadata (.library-index.json):
CODEBLOCK2

Fuzzy match query against metadata:

Match book?	Match topic?	→ Action
✅	✅	TOPIC (tiebreaker: future mixed searches)
✅

❌ | BOOK |
| ❌ | ✅ | TOPIC |
| ❌ | ❌ | CLARIFY (hard stop) |

Match rules:

• - Book: Query contains book title substring OR author name (case-insensitive)
• Topic: Query contains topic keyword (case-insensitive)

Examples

TOPIC wins (tiebreaker):

• - "Graeber debt finance" → matches both "Debt.epub" + "finance" → TOPIC: finance

BOOK only:

• - "Graeber hexagram 23" → matches "Debt.epub" only → BOOK: Debt.epub
• "I Ching moving lines" → matches "I Ching.epub" only → BOOK: I Ching.epub

TOPIC only:

• - "chaos magick sigils" → matches "chaos-magick" only → TOPIC: chaos-magick
• "mutual aid commons" → matches "anarchy" only → TOPIC: anarchy

CLARIFY (no match):

• - "philosophy" → no match → CLARIFY: "Search which topic or book?"
• "systems" → no match → CLARIFY: "Need more context - which area?"

Scope Types

1. 1. Topic scope: INLINECODE19

- Available topics: chaos-magick, finance, anarchy (check .topic-index.json)

1. 2. Book scope: INLINECODE20

- Requires exact filename (e.g., "Condensed Chaos.epub") - Use fuzzy matching: "Condensed" → "Condensed Chaos.epub"

---

Node 3-5: 👷 Call Wrapper

Execute wrapper script with inferred scope:

CODEBLOCK3

Arguments:

• - QUERY: User's search query (exact string)
• INLINECODE22: "topic" or "book"
• INLINECODE23: topic_id or book filename
• INLINECODE24: Number of results (default: 5)

Example calls:

CODEBLOCK4

---

Wrapper Exit Codes

The wrapper returns structured status via exit codes:

• - 0: Success (JSON results on stdout)
• 1: ERRORNOMETADATA (🤚 stop: tell user to run librarian index)
• 2: ERRORBROKEN (🤚 stop: system issue, report to Nicholas)
• 3: ERRORNO_RESULTS (🤚 stop: query returned 0 results)

Handle Each Error

Exit 1 (NO_METADATA):
CODEBLOCK5

Exit 2 (BROKEN):
CODEBLOCK6

Exit 3 (NO_RESULTS):

🤚 No results found for "[QUERY]"

Try:
- Broader terms (e.g., "debt" instead of "sovereign debt crisis")
- Different scope (search topic instead of single book?)
- Check spelling

---

Node 4: 🎤 Format Output

When wrapper returns success (exit 0), format the JSON results for the user.

JSON Structure

CODEBLOCK8

Formatting Rules

1. Synthesize answer (don't just list chunks)
- Read all results
- Extract key points
- Write coherent paragraph(s)

2. Cite sources with emojis
- 📕 = book citation
- Use location (page/paragraph) when available
- Format: INLINECODE26

3. Show similarity scores (optional, if useful)
- ⭐⭐⭐⭐⭐ (0.9-1.0) = Highly relevant
- ⭐⭐⭐⭐ (0.8-0.89) = Very relevant
- ⭐⭐⭐ (0.7-0.79) = Relevant

4. Keep original query context
- If user asked about debt, frame answer around debt
- Don't drift to tangential topics

Example Output

User: "What does Graeber say about the origins of money?"

Librarian:

Graeber argues that money did NOT originate from barter (the myth Adam Smith popularized). Instead, credit and debt systems came first — people kept track of obligations long before coins existed. 📕 [Debt: The First 5000 Years, p.21, ¶2]
He traces debt back to ancient Mesopotamia (~3500 BCE), where temple administrators recorded loans in cuneiform tablets. Money as we know it (coins) only appeared around 600 BCE in Lydia. 📕 [Debt, p.40, ¶5]
Key insight: Debt is older than money. Markets emerged from moral obligations, not rational barter. 📕 [Debt, p.89, ¶1]
Sources:

• - 📕 Debt: The First 5000 Years (David Graeber) - 3 passages
• Similarity: ⭐⭐⭐⭐⭐

---

Hard Stops (🤚 Honest Failures)

NEVER invent answers. If system fails, STOP and tell user exactly what's wrong.

When to Stop

1. 1. Metadata missing → Tell user to run INLINECODE27
2. Low confidence (<75%) → Ask clarifying question
3. System broken → Report error, don't guess
4. No results → Say "no results", suggest alternatives

Why Hard Stops Matter

From VISION.md: "Honest incompetence > false competence"

A broken skill that TELLS you it's broken is more trustworthy than one that invents plausible-sounding nonsense.

---

Installation & Setup

Requirements

• - Python 3.9+
• Dependencies: sentence-transformers, faiss-cpu, pypdf, INLINECODE31

Install

CODEBLOCK9

Index Your Library

CODEBLOCK10

---

Troubleshooting

"No metadata found"

• - Run index_library.py first
• Check books/.topic-index.json exists

"No results" but book exists

• - Check topic ID matches (e.g., "chaos-magick" not "chaos magick")
• Verify book is in correct topic folder
• Try broader query terms

"System broken"

• - Check Python dependencies: INLINECODE34
• Verify research.py syntax: INLINECODE35
• Check FAISS index integrity

---

References

Architecture:

• - Agentic Design Patterns (Andrew Ng, 2024) - Agentic workflows
• OpenClaw skill best practices - Protocol-driven skills

Sandwich pattern:

• - 🎤 Skill = Conversational I/O (trigger, infer, format, respond)
• 👷 Wrapper = Protocol enforcement (validate, build, check)
• ⚙️ Python = Heavy lifting (embeddings, search, ranking)

Why this works:

• - AI is good at: interpreting intent, formatting output, human communication
• AI is bad at: following syntax exactly, deterministic execution
• Wrapper hardens protocol: same query → same command → same behavior

---

Emoji Legend

• - 🎤 = Skill (AI conversational layer)
• 👷 = Wrapper (shell script protocol)
• ⚙️ = Python (research engine)
• 🤚 = Hard stop (honest failure)
• 📕 = Book citation
• ⭐ = Relevance score

---

Last updated: 2026-02-20 Epic: v0.15.0 Skill as Protocol