sxng

name: sxng description: "Search the web for current information, facts, docs, or research. Use when the user asks to search, look up, find, check, verify, or investigate anything — including current events, latest versions, comparisons, how-tos, troubleshooting, or any topic needing up-to-date info. Use deep search (--search-session) when comparing options, writing research reports, thoroughly investigating a topic, or when simple search is insufficient — watch for 'comprehensive comparison', 'detailed analysis', 'help me decide', 'which is better', 'write a report on', 'look into this thoroughly', 'cross-validate', or multi-dimensional questions. Also use when extracting web page content, or when a question cannot be answered from training data alone. Use this skill even if the user doesn't explicitly say 'search' — any question about current state of the world should trigger a web search."

SearXNG Web Search

Use sxng CLI to search the web. Results are automatically deduplicated (URL normalization + SimHash near-duplicate removal). CLI automatically filters out results with empty title or content.

Quick Reference

# Simple search
sxng <query>                                # Search (markdown output)
sxng --format json <query>                  # Search (JSON output)
sxng --queries "q1,q2,q3"                  # Multi-query with RRF fusion & dedup

# Deep search session
sxng --search-session new --owner "agent-1" --desc "topic" "query"
sxng --search-session <session> "more queries"

# Content extraction
sxng extract --urls "url1,url2"             # Extract from URLs
sxng extract --session <session>            # Extract session results
sxng extract --urls "url1" --obscura        # Extract with JS-rendering fallback

# Quality & iteration
sxng --search-session <session> --quality   # Assess result quality
sxng suggest-queries <session>              # Get query suggestions
sxng strategy-info <session>                # Check search stage
sxng recovery-analysis <session>            # Get recovery strategies
sxng session-report <session>               # Full session report

# Knowledge graph
sxng graph-preprocess <session>             # TF-IDF + co-occurrence analysis
sxng graph-add <session> --data '{...}'     # Add entities/edges
sxng graph-search <session> --keyword "x"   # Discover entities
sxng graph-search <session> --keyword "x" --limit 5  # Limit results
sxng graph-explore <session> --seed "x"     # View entity relations
sxng graph-drill <session> --seed "x" --relations "r1,r2"  # Follow relations
sxng graph-traverse <session> --path "p:chain_001"  # Traverse reasoning paths
sxng graph-obfuscate <session> --list       # List obfuscation candidates

# Utility
sxng session-list                           # List all sessions
sxng session-delete --older 24              # Delete old sessions
sxng --engines-list                         # List available engines
sxng --categories-list                      # List available categories
sxng --health                               # Check server status

Search Options

Extract

Extract full article content from web pages. Two-tier: Defuddle + linkedom (fast) → Obscura (JS rendering fallback).

sxng extract --urls "https://example.com/a,https://example.com/b"
sxng extract --session <session-name>
sxng extract --urls "https://spa-site.com" --obscura

--session (extract subcommand) vs --search-session (main search command) — different flags.

Deep Search

Deep search enables multi-round iterative research with quality assessment, recovery strategies, and knowledge graph navigation.

Read skills/sxng/references/SOP.md when doing any multi-result search — it covers source credibility tiers (white/grey/black list), cross-validation rules, and the Result Quality Filtering principle (keep liberally, filter conservatively). For deep search specifically, it also has L1/L2/L3 complexity levels, the full 8-phase workflow with Agent decision logic, and anti-patterns to avoid.

Quick Start

# Create session and search
sxng --search-session new --owner "agent-1" --desc "topic" "query"
# Output includes session path, use it for subsequent commands

# Extract content from results
sxng extract --session <session>

# Preprocess for entity discovery
sxng graph-preprocess <session>

# Add entities to knowledge graph
sxng graph-add <session> --data '{"entities":[...],"edges":[...]}'

# Assess quality
sxng --search-session <session> --quality

# Get suggestions if quality needs improvement
sxng suggest-queries <session>

# Continue searching with redundancy check
sxng "follow-up" --search-session <session> --redundancy warn

Session Management

Sessions stored under ~/sxng-cli/sessions/ by default.

sxng session-list                           # List all sessions
sxng session-delete <name>                  # Delete a session
sxng session-delete --older 24              # Delete sessions older than 24h

Quality Assessment

sxng --search-session <session> --quality

Returns 5 independent indicators: resultCount, contentDepth, entityRichness, sourceDiversity, novelty. Verdict: good / acceptable / poor. Based on verdict, use suggest-queries or recovery-analysis for next steps.

Knowledge Graph

Two layers:

• Structural (auto-built): query→result→domain nodes and edges
• Semantic (added by you via graph-add): entity nodes with custom relation edges

External search results: When using other search tools (tavily, exa, open-web-search, etc.) during a deep search session, inject their results into the graph via graph-add with the source field. This ensures the graph reflects all discoveries, not just sxng results. Result nodes carry a source field ("sxng" | "tavily" | "exa" | ...) — sxng-native results default to "sxng".

When adding edges, source/target must reference existing node IDs. Node ID prefixes:

Graph navigation commands: graph-search (discover), graph-explore (view relations), graph-drill (follow specific relations), graph-traverse (reasoning paths).

When to Stop

• Quality verdict is "good"
• Already 3+ search rounds
• New unique results < 3 per round

When to Use

Simple Search (sxng <query>): specific facts, API docs, error solutions, finding a URL.

Deep Search (--search-session): multi-dimensional comparison, cross-validation needed, research reports, or initial search reveals incomplete information.

See SOP for detailed L1/L2/L3 complexity guidelines.

Command Reference

Tips

• Default format: search & graph nav → md; analysis commands → json. Override with -f or --format.
• Results with empty title or content are automatically filtered out by CLI
• Use --time week/day for recent information
• Run sxng --health first if searches fail
• --session (extract subcommand) vs --search-session (main search command) — different flags
• Use --redundancy warn to avoid repeating similar queries
• Use --quality after each deep search round to decide whether to continue

Result Quality Filtering

When presenting search results to the user, apply lightweight quality judgment based on your own understanding. This is not a separate step — just a natural part of reading and outputting results.

Filter out results that clearly fall into these categories:

• The title and snippet are both unrelated to the query intent (not just missing keywords — semantically off-topic)
• The snippet is pure SEO stuffing: keyword repetition with no factual content
• The page is a known low-quality aggregator (mirrors, scrapers, parked domains) and adds no value beyond what other results already provide

Do NOT filter results that are:

• Partially relevant or tangentially related — they may provide useful context
• Written in a different style or tone than expected — relevance matters more than presentation
• Missing some query terms but still on-topic — semantic match > keyword match
• From lesser-known sources — small blogs and forums often have the best answers

Principle: keep liberally, filter conservatively. When uncertain about a result's relevance, always keep it. It is better to show a slightly noisy result than to hide a useful one.