Skip to content

CLI Reference

Basic Usage

truth-eval <document> [options]

Options

Option Description Default
--root-path, -r Root directory for filesystem search None
--output, -o Output file (auto-detects format from extension) None
--model, -m Model to use (can specify multiple) .env or gpt-4o
--confidence, -c Confidence threshold .env or 0.7
--web-search / --no-web-search Enable/disable web search .env or True
--human-review / --no-human-review Enable human-in-the-loop .env or False
--mode Verification mode: external, internal, both external

Examples

Basic

truth-eval README.md

With Filesystem Context

truth-eval README.md --root-path .

Multi-Model

truth-eval README.md \
  --model gpt-4o \
  --model gpt-4o-mini \
  --model claude-sonnet-4-5

Fireworks AI

export FIREWORKS_API_KEY="fw_..."
truth-eval README.md --model accounts/fireworks/models/llama-v3-8b-instruct

Save Report

truth-eval README.md --output report.json
truth-eval README.md -o report.md

Strict Verification

truth-eval README.md --confidence 0.9

Web search is enabled by default. To run without web search, use filesystem evidence only:

truth-eval README.md --root-path . --mode internal

Human Review

truth-eval README.md --human-review

Pauses for low-confidence claims: 

Claim: Python requires 3.11+
Proposed: NOT_ENOUGH_INFO (40%)
Approve? (approve/correct:SUPPORTS/skip)

Output Formats

Terminal (Default)

Rich tables and panels:

╭────────────────────── Evaluation Summary ──────────────────────╮
│ Grade: A+                                                      │
│ Confidence: 91.7%                                              │
╰────────────────────────────────────────────────────────────────╯

truth-eval README.md -o report.md

Generates a readable report with: - Executive summary - Detailed claim-by-claim analysis - Evidence sources - Model votes

JSON

truth-eval README.md -o report.json

{
  "overall_grade": "A+",
  "overall_confidence": 0.917,
  "claims": [...],
  "verifications": [...],
  "statistics": {
    "total_claims": 4,
    "supported": 3,
    "refuted": 0,
    "not_enough_info": 1
  }
}

HTML

truth-eval README.md -o report.html

Self-contained HTML with styling.

Exit Codes

Code Meaning
0 Success
1 Error (file not found, API error, etc.)

Configuration File

The CLI reads .env automatically. CLI flags override .env values.

# .env
TRUTH_VERIFICATION_MODELS=["gpt-4o","claude-sonnet-4-5"]
TRUTH_CONFIDENCE_THRESHOLD=0.8
TRUTH_ENABLE_WEB_SEARCH=false

# Uses .env defaults
truth-eval README.md

# Overrides only the models
truth-eval README.md --model gpt-4o

Flag Precedence

  1. CLI flag (explicit) → wins
  2. .env file value
  3. Built-in default

Boolean Flags

Use --flag or --no-flag:

# Explicitly disable web search (overrides .env)
truth-eval README.md --no-web-search

# Explicitly enable human review (overrides .env)
truth-eval README.md --human-review

Environment Variables

Override defaults inline:

TRUTH_CLAIM_EXTRACTION_MODEL=gpt-4o-mini \
TRUTH_CONFIDENCE_THRESHOLD=0.6 \
truth-eval README.md