terraphim/terraphim-ai/pre-llm-validate

Pre-LLM Validation

Use this skill to validate input before sending to LLMs. Ensures semantic coherence and domain relevance using the local knowledge graph.

When to Use

• Before making LLM API calls with user-provided context
• When validating that input contains relevant domain terms
• To check semantic coherence of multi-concept queries
• Pre-flight validation to reduce wasted LLM tokens

Validation Steps

1. Check Semantic Connectivity

# Validate that input terms are semantically connected
terraphim-agent validate --connectivity --json "your input text"

Output interpretation:

• connected: true - Input is semantically coherent, proceed with LLM call
• connected: false - Input spans unrelated concepts, consider warning user

2. Check Domain Coverage

# Find matched domain terms
terraphim-agent validate --json "your input text"

Output interpretation:

• High match count - Input is domain-relevant
• Low/zero matches - Input may be off-topic or use non-standard terminology

3. Suggest Corrections (Optional)

If terms aren't found, suggest alternatives:

# Get fuzzy suggestions for potential typos
terraphim-agent suggest --fuzzy "unclear term" --threshold 0.6

Workflow Example

# 1. Validate connectivity
RESULT=$(terraphim-agent validate --connectivity --json "$INPUT")
CONNECTED=$(echo "$RESULT" | jq -r '.connected')
 
if [ "$CONNECTED" = "false" ]; then
    echo "Warning: Input spans unrelated concepts"
    TERMS=$(echo "$RESULT" | jq -r '.matched_terms | join(", ")')
    echo "Matched terms: $TERMS"
fi
 
# 2. Check domain coverage
MATCHES=$(terraphim-agent validate --json "$INPUT" | jq '.matched_count')
 
if [ "$MATCHES" -lt 2 ]; then
    echo "Warning: Low domain coverage ($MATCHES terms)"
fi
 
# 3. Proceed with LLM call if validation passes

Integration with Hooks

This skill can be automated via the pre-tool-use hook:

terraphim-agent hook --hook-type pre-tool-use --input "$JSON"

Configuration

• --role - Specify role for domain-specific validation
• --json - Machine-readable output for scripting
• --threshold - Similarity threshold for fuzzy matching (default: 0.6)

Best Practices

1. Advisory Mode: Use validation results as warnings, not blockers
2. Role Selection: Match role to the domain of the query
3. Threshold Tuning: Adjust fuzzy threshold based on domain vocabulary
4. Caching: Cache validation results for repeated queries