Research

Research and Source Ingestion

This skill provides comprehensive guidance for autonomous research execution, source evaluation, and corpus ingestion.

Source Reliability Evaluation Framework

Evaluate every source using these criteria before ingestion. Reliability affects how confidently claims can be made.

Reliability Tiers

Reliability Scoring Checklist

For each source, evaluate:

Authority (0-3 points):

• Author is subject matter expert (1 pt)
• Published by reputable institution (1 pt)
• Peer-reviewed or fact-checked (1 pt)

Evidence (0-3 points):

• Cites primary sources (1 pt)
• Includes references/bibliography (1 pt)
• Provides specific details, not generalizations (1 pt)

Recency (0-2 points):

• Published within relevant timeframe for topic (1 pt)
• Updated or confirmed still accurate (1 pt)

Bias/Objectivity (0-2 points):

• Acknowledges limitations or counterarguments (1 pt)
• Not promotional or heavily biased (1 pt)

Total Score → Reliability Tier:

• 8-10 points: High
• 5-7 points: Medium
• 2-4 points: Low
• 0-1 points: Very Low

Source Type Definitions

Primary Sources (reliability: High):

• Original documents (letters, diaries, official records)
• Eyewitness accounts (interviews, memoirs from participants)
• Raw data (statistics, research datasets)
• Artifacts (photographs, objects, recordings)

Secondary Sources (reliability: Medium to High):

• Scholarly analysis of primary sources
• Historical or academic syntheses
• Biographies by credible historians
• Well-researched journalism

Tertiary Sources (reliability: Low):

• Encyclopedias (Wikipedia, Britannica)
• Textbooks (introductory overviews)
• Almanacs and fact books
• Bibliographies

Web Sources (reliability: Varies):

• Evaluate individually using checklist
• Domain matters: .edu, .gov often more reliable than .com
• Check "About" page for authorship and credentials
• Verify claims against higher-tier sources

Web Search Strategies

Strategy 1: Targeted Keyword Search

When to use: You know the specific fact, person, or event you need.

Technique:

"exact phrase" + broad context + qualifier

Examples:
"SOE Beaulieu training" + 1942 + protocol
"Lyon resistance network" + 1943 + structure
"wireless operator" + occupied France + procedure

Query operators:

• "exact phrase" — forces exact match
• site:domain.com — restrict to domain
• filetype:pdf — find PDFs (often academic papers)
• -exclude — remove unwanted terms
• OR — alternatives (capitalize for boolean)

Strategy 2: Progressive Refinement

When to use: Broad topic, need to narrow down.

Technique:

1. Start broad: "SOE training"
2. Review top results for specific terms
3. Refine: "SOE training Beaulieu wireless"
4. Iterate until hitting primary sources or academic work

Strategy 3: Reverse Citation Chase

When to use: Found one good source, need more.

Technique:

1. Find one high-quality source
2. Extract author names, key terms, referenced works
3. Search for those authors' other publications
4. Search for works that cite this source (Google Scholar: "Cited by")

Strategy 4: Academic Database Search

When to use: Need scholarly rigor for nonfiction or historical fiction.

Databases to use:

• Google Scholar (free, broad coverage)
• JSTOR (subscription, humanities/social sciences)
• PubMed (free, medical/scientific)
• Archive.org (free, historical documents)

Technique:

• Use academic keywords (avoid colloquialisms)
• Filter by date range
• Sort by citation count for influential works

Strategy 5: Primary Source Discovery

When to use: Fiction requiring historical accuracy, nonfiction requiring evidence.

Resources:

• National archives (e.g., UK National Archives, US National Archives)
• University special collections
• Digital humanities projects
• Museum databases

Technique:

• Search "[topic] primary sources"
• Search "[topic] archive collection"
• Look for digitized documents, oral histories

Ingestion Workflow

Follow this workflow for each source:

Step 1: Chunk

Break source into semantically coherent pieces using LLM-based chunking.

Chunking strategy: semantic (not fixed token windows)

Process:

1. LLM identifies natural breakpoints (topic shifts, scene changes, paragraph boundaries)
2. Max tokens per chunk: 1024 (configurable via bookstrap.config.json)
3. Overlap between chunks: 128 tokens (preserves context)

Why semantic chunking?

• Preserves meaning (doesn't split mid-thought)
• Better for retrieval (chunks are topically coherent)
• LLM can identify section headers, topic transitions

Implementation:

• Use host framework's LLM (Claude, Gemini, etc.)
• Prompt: "Identify natural breakpoints for chunking this document. Return byte offsets."

Step 2: Embed

Generate vector embeddings for each chunk.

Process:

1. Send chunk text to embedding provider (configured in bookstrap.config.json)
2. Store embedding in source.embedding or section.embedding field
3. SurrealDB native vector type: array<float>

Embedding providers (via config):

• Gemini: text-embedding-004 (768 dims)
• OpenAI: text-embedding-3-small (1536 dims)
• Ollama: nomic-embed-text (768 dims, local)
• LM Studio: Local embeddings (768 dims)

Dimensions must match config setting for vector similarity queries to work.

Step 3: Extract

Use LLM to extract entities and relationships from each chunk.

Entities to extract:

• Characters (fiction): name, description, status (alive/dead)
• Locations: name, description, introduced (bool)
• Events: name, description, sequence, date (if available)
• Concepts (nonfiction): name, description

Extraction prompt pattern:

Analyze this text and extract:
1. People mentioned (name, role, description)
2. Locations (name, description)
3. Events (name, description, date if mentioned)
4. Key concepts or themes

For each entity, also identify:
- Relationships (who knows whom, what relates to what)
- Timeline information (when did this happen?)

Text:
[chunk content]

Store entities:

-- Example: Create character
CREATE character SET
  name = "Anna",
  description = "SOE wireless operator, recruited 1942",
  status = "alive",
  embedding = $embedding_vector;

Step 4: Relate

Create graph relationships between entities.

Relationship types:

• appears_in: character → section
• located_in: section → location
• cites: section → source
• supports: source → concept
• precedes: event → event (chronological)
• follows: event → event (inverse of precedes)
• knows: character → character
• related_to: concept → concept

Storage pattern:

-- Link character to section
RELATE character:anna->appears_in->section:ch3_sec2;

-- Link section to cited source
RELATE section:ch3_sec2->cites->source:soe_manual_1942;

-- Link source to supported concept
RELATE source:soe_manual_1942->supports->concept:wireless_protocols;

-- Timeline ordering
RELATE event:training_begins->precedes->event:deployment;

Why relationships matter:

• Enable graph queries ("What do we know about Anna?")
• Enforce consistency (can't mention dead character as alive)
• Support citation tracking (every claim → source)

Step 5: Quality Check

Verify ingestion quality before marking task complete.

Thresholds (reject source if fails):

• At least 1 entity extracted
• Embedding successfully generated (vector not null)
• Source reliability scored (not "unknown")
• Source metadata complete (title, URL, source_type, ingested_at)

Validation queries:

-- Check if entities were created
SELECT count() FROM character WHERE id IN (SELECT <-appears_in<-section<-cites<-source WHERE id = $source_id);

-- Check embedding exists
SELECT embedding FROM source WHERE id = $source_id AND embedding IS NOT NONE;

If quality check fails:

• Log warning with source ID
• Flag for manual review
• Do not mark knowledge_gap as resolved

Ingestion Storage Pattern

Full ingestion creates these database records:

-- 1. Create source record
CREATE source SET
  title = "SOE Training Manual 1942",
  content = $full_text,
  embedding = $doc_embedding,
  url = "https://example.com/soe-manual",
  source_type = "primary",
  reliability = "high",
  ingested_at = time::now(),
  ingested_during = "research";  -- or "bootstrap", "writing"

-- 2. Create entities found in source
CREATE character SET
  name = "Anna",
  description = "SOE wireless operator",
  embedding = $entity_embedding,
  introduced_in = section:none;  -- will link when writing

-- 3. Create relationships
RELATE source:soe_manual->supports->concept:wireless_training;

-- 4. Update timeline if dates found
CREATE event SET
  name = "Beaulieu training begins",
  description = "Anna starts SOE wireless operator training",
  sequence = 5,
  date = "1942-08-15T00:00:00Z";

-- 5. Mark knowledge gap resolved
UPDATE knowledge_gap:gap_12 SET
  resolved = true,
  resolved_by = source:soe_manual;

Provider Configuration

Research providers are configured in bookstrap.config.json:

Web Search Providers

{
  "research": {
    "provider": "tavily",  // or "brave", "serper", "google"
    "api_key_env": "TAVILY_API_KEY",
    "rate_limit": {
      "requests_per_minute": 10
    },
    "blocked_domains": ["example-spam.com"],
    "allowed_domains": [],  // if set, only fetch from these
    "max_sources_per_task": 5
  }
}

Provider selection:

• Tavily: Best for research-focused queries, returns high-quality sources
• Brave: Privacy-focused, good general search
• Serper: Google results via API
• Google: Direct Google API (requires custom search setup)

Embedding Providers

{
  "embeddings": {
    "provider": "gemini",  // or "openai", "ollama", "lmstudio"
    "model": "text-embedding-004",
    "dimensions": 768,
    "api_key_env": "GEMINI_API_KEY"
  }
}

Dimensions must match across all embeddings in a database. Cannot mix 768-dim and 1536-dim vectors.

Research Task Execution Pattern

When executing a research task:

1. Load task: SELECT * FROM knowledge_gap WHERE id = $task_id
2. Web search: Use configured provider with task.question
3. Evaluate sources: Apply reliability framework to each result
4. Select sources: Pick top max_sources_per_task by reliability score
5. Ingest each source:
   • Fetch content
   • Chunk semantically
   • Generate embeddings
   • Extract entities
   • Create relationships
6. Quality check: Verify thresholds met
7. Mark resolved: UPDATE knowledge_gap SET resolved=true, resolved_by=$source_id
8. Commit: git commit -m "[bookstrap] research: [task description]"

Example Research Workflow

# Task: Research "SOE wireless operator training protocols"

1. Web search query: "SOE wireless operator training" + Beaulieu + 1942 + protocol
2. Results:
   - nationalarchives.gov.uk/soe/training → High reliability (primary source)
   - wikipedia.org/SOE → Low reliability (tertiary, use for leads only)
   - soe-history-blog.com → Very Low (personal blog, skip)

3. Select: nationalarchives.gov.uk document
4. Fetch content (text extraction from HTML/PDF)
5. Chunk: LLM finds 12 semantic chunks
6. Embed: Generate 768-dim vectors via Gemini
7. Extract:
   - Event: "Wireless training begins at Beaulieu, August 1942"
   - Concept: "Morse code proficiency required 20 WPM"
   - Concept: "Encryption protocols for field operations"
8. Relate:
   - RELATE source->supports->concept:morse_proficiency
   - CREATE event SET name="Training begins", date="1942-08-01"
9. Quality check: ✓ 3 entities, ✓ embedding, ✓ reliability=high
10. Mark resolved: UPDATE knowledge_gap:gap_12 SET resolved=true
11. Commit: git commit -m "[bookstrap] research: SOE wireless training protocols"

Handling Research Failures

If no sources found:

• Log failure reason
• Leave knowledge_gap unresolved
• Report to user: "Could not find reliable sources for [topic]"
• Suggest: broaden search, try alternative keywords, or mark as low-priority

If only low-reliability sources found:

• Ingest with reliability = "low"
• Flag for user review
• Note in commit message: "Low-reliability sources only"
• Consider marking gap as "partially resolved"

If ingestion fails:

• Check URL accessibility
• Verify API keys configured
• Check rate limits
• Retry with backoff
• If persistent failure: skip source, try next result

Output and Reporting

After completing a research task:

## Research Task Complete: [topic]

**Sources ingested**: [count]
- High reliability: [count]
- Medium reliability: [count]

**Entities extracted**:
- Characters: [count]
- Locations: [count]
- Events: [count]
- Concepts: [count]

**Knowledge gap resolved**: [gap ID]

**Next steps**:
- Continue with next research task
- OR return to writing (if all gaps resolved)

Quality Thresholds

Reject sources that fail these criteria:

Supporting Files

This skill references additional resources:

• web-search.md — Detailed search strategies by genre/topic
• extraction.md — Entity extraction patterns and prompts
• scripts/chunk.py — Semantic chunking implementation
• scripts/embed.py — Embedding generation utilities

Load these files when needed for specific sub-tasks.