FAQ

FAQ

Common questions about rvrb-verify.

Provider setup

Which provider should I use?

Qwen (default) — Best for Qwen models, free tier available via Alibaba Cloud.

OpenAI — Use if you have an OpenAI API key and prefer GPT models.

Local — Use Ollama or vLLM for offline operation. No API key needed.

How do I get a Qwen API key?

1. Go to Alibaba Cloud Model Studio
2. Sign up for a free account (Singapore region)
3. Activate Model Studio
4. Generate an API key
5. Set export DASHSCOPE_API_KEY="sk-..."

Free tier: 1M tokens per model, 90-day validity.

Can I use different models for search and judge?

Yes:

from rvrb_verify import verify, get_provider

search = get_provider(model="qwen3-coder-plus")  # Fast model for search
judge = get_provider(model="qwen3.7-plus")       # Smart model for judge

verdict = verify("Claim", search_provider=search, judge_provider=judge)

Strategy selection

Which strategy should I use?

Claim type	Strategy
General facts, news	fact-check
Legal, regulatory	legal
Scientific, academic	research

Can I create custom strategies?

Yes. See Development for details on adding custom strategies.

What tools does each strategy use?

• fact-check: web search, news search
• legal: statute search, case law search
• research: paper search, arXiv search

Confidence scores

What does confidence mean?

Confidence is a float from 0.0 to 1.0 representing the judge model's certainty in the verdict.

• 0.9+: High confidence
• 0.7-0.9: Medium confidence
• <0.7: Low confidence, may be inconclusive

Why is confidence low?

Low confidence can indicate:

• Ambiguous or subjective claim
• Insufficient evidence found
• Conflicting sources
• Complex or nuanced topic

Can I set a confidence threshold?

Not currently. You can filter results programmatically:

verdict = verify("Claim")
if verdict.confidence < 0.7:
    print("Low confidence, manual review needed")

Offline usage

Can I use rvrb-verify offline?

Yes, with a local provider:

# Install Ollama
# Pull a model
ollama pull llama3

# Use local provider
rvrb-verify "Claim" --provider local

Do the tools work offline?

No. The tools (web search, news search, etc.) require internet access. For offline use, implement a custom ToolGateway with local data sources.

Token costs

How many tokens does verification use?

Approximate usage per verification:

• Search phase: 500-2000 tokens (depends on tools used)
• Judge phase: 500-1000 tokens
• Total: 1000-3000 tokens per verification

How can I reduce costs?

• Use faster/cheaper models for search phase
• Limit evidence collection
• Use local provider (no API costs)

Integration

Can I use rvrb-verify with other satellites?

Yes. Common pipelines:

# Transcribe audio → verify claims
rvrb-transcribe meeting.mp3 | rvrb-verify

# Verify → format results
rvrb-verify "Claim" --json | rvrb-transform "Format as markdown table"

Can I use rvrb-verify in my application?

Yes. Install via pip and use the Python API:

from rvrb_verify import verify

verdict = verify("Claim")

Troubleshooting

Error: "DASHSCOPE_API_KEY is not set"

Set the environment variable:

export DASHSCOPE_API_KEY="sk-..."

Or use a different provider:

rvrb-verify "Claim" --provider openai

Error: "Unknown strategy: xyz"

Check available strategies:

python -c "from rvrb_verify import list_strategies; print(list_strategies())"

Available: fact-check, legal, research.

Error: "Model not found"

Check that the model ID is correct:

rvrb-verify "Claim" --model qwen3-coder-plus

For Qwen models, see the model catalog.

Verification returns "inconclusive"

The judge model couldn't determine the verdict with high confidence. This can happen with:

• Ambiguous claims
• Insufficient evidence
• Conflicting sources

Try:

• Rephrasing the claim
• Using a different strategy
• Using a more capable model

Tools return empty results

The MockToolGateway returns mock data. For real tool execution, implement a RealToolGateway (see Development).

Performance

How long does verification take?

Typical verification time:

• Search phase: 2-10 seconds (depends on tools)
• Judge phase: 1-3 seconds
• Total: 3-13 seconds

Can I speed up verification?

• Use faster models (e.g., qwen3-coder-flash for search)
• Reduce evidence collection
• Use local provider (no network latency)

License

Apache-2.0 — same as the reverberage ecosystem.