answer-engine

name

answer-engine

description

Research-focused query handling with multi-source synthesis, citations, and Obsidian persistence. Like a self-hosted Perplexity/Vane but CLI-native. Best for quick-to-medium lookups using Kagi. Use when the user asks factual questions, needs citations, or wants a direct answer — not a full research report (use deep-research) or social sentiment (use last30days). Triggers on: research, look into, what's the latest on, compare, explain, investigate.

Answer Engine Skill

Purpose

Research-focused query handling with multi-source synthesis, citations, and Obsidian persistence. Like a self-hosted Perplexity/Vane, but CLI-native.

Trigger

User asks questions requiring research, verification, or multi-source synthesis. Best for quick-to-medium lookups using Kagi as the search backend. For systematic 4-phase deep research reports, use research/deep-research instead. For social-signal / "what people are saying" research, use last30days instead. See research/deep-research/references/research-skill-routing.md for the full routing table.

Search Backend Stack

1. Kagi tools plugin — Use kagi_search (structured results), kagi_quick (factual answers with references), or kagi_assistant (open-ended research conversations). These are Hermes-native tools, no terminal required.
   • kagi_search — Best for finding sources with filters (region, time, lens)
   • kagi_quick — Best for single-answer factual queries with citations
   • kagi_assistant — Best for deep research with thread continuation
   • kagi_ask_page — Best for understanding a specific URL's content
   • Note: kagi_fastgpt is deprecated — use kagi_quick instead
2. SearXNG - Secondary, privacy-focused diversity (currently degraded — returns 0 results as of 2026-05; keep trying but don't rely on it)
   • Instance: https://searxng.snakepit.us
   • Query format: https://searxng.snakepit.us/search?q={query}&format=json
3. web_search - Not available in current toolset (stale reference, do not use)

Workflow

1. Classify Query

Determine intent:

• Factual: Direct answer exists (e.g., "what is X")
• Research: Requires synthesis (e.g., "compare X vs Y", "best X for Y")
• Technical: Code, APIs, debugging
• File-based: References uploaded documents
• Exploratory: Open-ended, emerging topics (requires broader net)

2. Execute Searches (Parallel when possible)

Minimum source targets:

• Factual: 3-5 sources
• Research: 5-10 sources
• Exploratory: 8-15 sources

Search stack (as of 2026-05):

• Call kagi_quick <query> for quick factual answers with citations (replaces deprecated kagi fastgpt)
• Call kagi_search <query> for detailed search results with region/time/lens filters
• Query SearXNG JSON API for diverse/independent sources (currently degraded, 0 results as of 2026-05)
• For complex topics: delegate_task with batch mode (split into sub-questions)
• Note: web_search tool does NOT exist in current toolset — do not reference it

Query expansion: Generate 2-3 related search queries to capture adjacent angles:

• Example: "solid-state battery breakthroughs" → also search "quantum scape 2026", "toyota solid state battery timeline", "lithium metal anode challenges"

3. Track Sources with Quality Signals

For each source, record:

• Source URL
• Source title
• Publication type (academic, news, blog, forum, official docs)
• Date/timeliness
• Credibility indicators (peer-reviewed, primary source, expert author, etc.)
• Citation number [1], [2], etc.

4. Semantic File Search (if applicable)

If user references files or asks about prior research:

• Use NanoGPT API with gemini-embedding-2-preview
• Embed query, search Obsidian Research folder
• Include relevant excerpts as sources
• Note connections to prior research explicitly

5. Deep Synthesis (Multi-Pass)

Pass 1 - Extraction: Pull key facts, claims, data points from each source Pass 2 - Clustering: Group related findings into thematic buckets Pass 3 - Tension Mapping: Identify where sources agree, disagree, or fill gaps Pass 4 - Narrative: Build coherent answer with logical flow

Output structure for research-grade answers:

• Executive Summary (2-4 sentences, high-level answer)
• Key Findings (bulleted, with citations)
• Detailed Analysis (paragraphs with synthesis, not just source regurgitation)
• Contradictions/Uncertainties (explicitly note disagreements or gaps)
• Methodology Notes (what was searched, limitations)

6. Format Output

# {Query as Title}

## Executive Summary

{2-4 sentence high-level answer}

## Key Findings

- {Finding 1} [1][2]
- {Finding 2} [3]
- {Finding 3} [4][5]

## Detailed Analysis

{Multi-paragraph synthesis with inline citations}

## Contradictions & Uncertainties

{Note where sources disagree or information is incomplete}

## References

[1] Source Title - URL (publication type, date)
[2] Source Title - URL (publication type, date)

---
*Research saved to Obsidian: {filename}*

7. Save to Obsidian

• Path: ~/Documents/Obsidian Vault/Research/
• Filename: {YYYY-MM-DD}-{slugified-query}.md
• Include full research + metadata (timestamp, sources, search terms, query expansions)
• Tag with relevant topics for future retrieval

Note: After saving, the research file is automatically ingested into the llm-wiki knowledge base via the wiki-ingestion skill. This creates a query/ page in the wiki, auto-detects and creates entity pages for notable mentions, and updates the index. See research/wiki-ingestion/INTEGRATIONS.md for details.

Contact & Person Lookup

When the user needs to find someone's email, phone, or contact info at an organization (university, company, etc.), see references/contact-lookup.md for the directory scraping technique, email naming convention inference, and fallback strategies.

Search Mode

Deep Research (default):

• Factual queries: 5-8 sources minimum
• Research/synthesis queries: 8-15 sources minimum
• Exploratory queries: 10-20 sources minimum
• Kagi primary, SearXNG secondary when available (currently degraded)
• Deep synthesis with multi-pass analysis
• Target: 3-5 minutes for thorough coverage (quality over speed)
• Always include: executive summary, key findings, detailed analysis, uncertainties

Quick Mode (only when user explicitly asks for "quick" or "brief"):

• 3-5 sources
• Single-pass synthesis
• Target: <90 seconds

Citation Rules

• Number sequentially [1], [2], [3]
• Each number = one URL
• Order by relevance, not discovery order
• Include in References section at end
• Reference format: [N] Title - URL (publication type, date if available)
• Source type tags: Use parentheticals like (academic), (news), (official docs), (blog), (forum), (primary source)
• Multiple citations per claim: When a claim is supported by multiple sources, cite all: "Solid-state batteries promise 2-3x energy density [1][3][7]"
• Contradiction citations: When sources disagree, cite each side: "Toyota targets 2027 launch [2], while QuantumScape claims 2025 [5]"

File Embeddings

NanoGPT API config:

• Model: gemini-embedding-2-preview
• Use for: Semantic search over Research folder
• Embed new research files after saving
• Fallback: If API is unreachable, skill uses keyword-based search_files instead

Tools Used

• kagi_search — Primary search (structured JSON results), available as a Hermes tool
• kagi_quick — Factual answers with references (replaces deprecated kagi fastgpt)
• kagi_assistant — Deep research conversations with thread continuation
• kagi_ask_page — Page-specific questions with citations
• terminal — SearXNG JSON queries (degraded), Obsidian operations
• delegate_task — Parallel research (complex queries) — also useful for browser-based search when JS rendering needed
• execute_code — Embedding API calls, text processing, citation tracking
• write_file — Obsidian output
• search_files — Keyword-based file search (fallback if embeddings unavailable)
• Note: web_search is NOT available — use kagi_search instead

Wiki Integration

After research is saved to Obsidian, it is automatically ingested into the llm-wiki knowledge base (see research/wiki-ingestion skill). This means:

• The research becomes a permanent, cross-referenced wiki page (not a disposable file)
• Entities (tools, people, companies, models) mentioned are auto-detected
• New entity pages are created for previously unknown entities
• Existing entity pages are updated with references to this research
• The wiki index and log are updated automatically

This creates a compounding knowledge base where each research session builds on previous ones through cross-links, rather than accumulating isolated files.

Pitfalls

• SearXNG may rate-limit — add 2s delay between queries
• NanoGPT API key in ~/.config/nanogpt/.env
• Obsidian path has space — quote paths in shell commands
• Don't over-cite: 1 citation per distinct claim, not every sentence
• Country/comparison statistics: Comparison aggregator sites (versus.com, mylifeelsewhere.com, georank.org, countryeconomy.com) frequently have outdated, misread, or just wrong data. Examples: one claimed Cuba's obesity rate was 2.6% (actually 24.6% per CIA Factbook), another claimed US literacy was 86% (UNESCO has no recent US data; CIA Factbook says ~99%). Always verify against primary sources: CIA World Factbook, World Bank Open Data, WHO, UNESCO Institute for Statistics, NCES (for US education stats). If a comparison site's number seems surprising or dramatic, it's probably wrong — verify before including it.

Usage Examples

See references/examples.md for detailed walkthroughs of:

• Simple Factual Query
• Comparative Research (Deep Dive)
• Exploratory Research (Emerging Topic)
• File-Referenced Query

Script Reference

searxng_search.py

python scripts/searxng_search.py "search query"
# Returns: JSON array of {title, url, content, engine, score}

citation_tracker.py

python scripts/citation_tracker.py add "Source Title" "https://url.com" "publication_type" "date"
python scripts/citation_tracker.py generate
python scripts/citation_tracker.py clear
python scripts/citation_tracker.py export  # Export as markdown references section

save_to_obsidian.py

python scripts/save_to_obsidian.py "Query text" "Full markdown content" "tags"
# Returns: Full filepath where saved

embedding_search.py

# Build index (run once, or when Research folder changes significantly)
python scripts/embedding_search.py rebuild

# Search for similar documents
python scripts/embedding_search.py search "your query"

# Embed a new file (auto-called after saving research)
python scripts/embedding_search.py embed "/path/to/file.md"

query_expander.py

python scripts/query_expander.py "solid-state battery breakthroughs"
# Returns: JSON array of related queries:
# ["quantum scape 2026", "toyota solid state battery timeline", "lithium metal anode challenges"]

source_evaluator.py

python scripts/source_evaluator.py "https://url.com"
# Returns: {publication_type, credibility_score, date, author_info, is_primary_source}
# Credibility signals: domain age, citation count, author credentials, peer-review status

Credential Setup

# NanoGPT API key for embeddings (already configured)
# Located at: ~/.config/nanogpt/.env

# Kagi tools plugin is already installed and configured
# Available tools: kagi_search, kagi_quick, kagi_assistant, kagi_ask_page,
#                  kagi_translate, kagi_summarize, kagi_news, kagi_smallweb
# Session token is in env (KAGI_SESSION_TOKEN)

Status

• ✅ Kagi tools plugin (kagi_search, kagi_quick, kagi_assistant, etc.): Working, primary search backend
• ⚠️ SearXNG: Connected but returning 0 results (snakepit.us, degraded since ~2026-05)
• ❌ web_search: Not available in current toolset
• ✅ Obsidian Research folder: Exists
• ✅ NanoGPT API key: Configured
• ✅ Embeddings: Working (text-embedding-3-small, 12 files indexed)
• ✅ Deep Research Mode: Active (default)
• ✅ Query Expansion: Available via query_expander.py
• ✅ Source Evaluation: Available via source_evaluator.py