Skip to content

faq.mdx

Latest commit

 

History

History
363 lines (260 loc) · 8.36 KB

faq.mdx

File metadata and controls

363 lines (260 loc) · 8.36 KB
Raw
sidebar_position 10
title Frequently asked questions
description Get answers to common questions about QWED verification. Learn how it differs from RAG, fine-tuning, and guardrails for deterministic LLM output validation.

General

What is QWED?

QWED is a deterministic verification protocol for Large Language Models (LLMs). It treats LLMs as "untrusted translators" and verifies their outputs using formal methods (SymPy, Z3, AST, SQLGlot).

How is QWED different from other LLM tools?

Feature QWED RAG Fine-tuning Guardrails
Deterministic ✅ Yes ❌ No ❌ No ❌ No
Mathematically proven ✅ Yes ❌ No ❌ No ⚠️ Limited
No training required ✅ Yes ⚠️ Needs docs ❌ No ✅ Yes
Works offline ❌ API-based ✅ Yes ✅ Yes ✅ Yes

QWED complements these tools — use RAG for knowledge, QWED for verification.

Do I need to run qwed init every time?

No. Once initialized, QWED reads your provider credentials and configuration from the .env file that qwed init creates. You only need to re-run it when you want to switch LLM providers or rotate API keys. Running qwed init again merges new values into your existing .env — it does not overwrite settings you don't change.

Integration

Do I need to run a backend server?

Yes. QWED requires a backend server with your LLM API keys configured.

Architecture:

Your app → SDK → Backend server (you run) → LLM (your key) → Verifiers

Setup:

# Step 1: Configure your LLM
cp .env.example .env
echo "ANTHROPIC_API_KEY=sk-ant-..." >> .env

# Step 2: Run backend
python -m qwed_api

# Step 3: Use SDK
python your_app.py

See Getting started for full setup.

Do I need to call my LLM first?

No. This is a common mistake.

❌ Wrong:

llm_result = openai.chat(...)  # Don't do this
qwed.verify(llm_result)  # Too late

✅ Correct:

# Let QWED backend handle LLM call
result = qwed.verify("your question")

The backend server (that you run) calls the LLM using your API key.

What LLM providers does QWED support?

You configure your LLM provider in the backend's .env file:

Supported providers:

  • OpenAI (direct API)
  • Anthropic Claude
  • Azure OpenAI
  • AWS Bedrock
  • Google Gemini

Example .env:

ACTIVE_PROVIDER=anthropic
ANTHROPIC_API_KEY=sk-ant-...

See LLM configuration for all providers.

Can I use my own LLM API key?

Yes. You must use your own LLM API key. QWED is open source — you run the backend server with your credentials.

You provide:

  • Your own LLM API key (in .env)
  • Your own backend server (run locally)

You control:

  • Which LLM provider to use
  • All your data and keys

Costs

How much does QWED cost?

Open source: free.

  • You pay only for your own LLM API usage
  • No QWED subscription needed
  • Run backend server yourself

Costs you pay:

  • Your LLM provider (OpenAI, Anthropic, etc)
  • Your hosting (if deploying backend)

Example: If using Anthropic Claude:

  • Input: $3 per million tokens
  • Output: $15 per million tokens
  • (See your LLM provider's pricing)

What are my rate limits?

QWED itself: no rate limits. It's open source.

Your LLM provider: Check their limits:

  • OpenAI: Tier-based (see dashboard)
  • Anthropic: Based on plan
  • Azure: Based on deployment

If you hit your LLM provider's rate limit, implement retries:

import time

try:
    result = qwed.verify(query)
except Exception as e:
    if "rate_limit" in str(e):
        time.sleep(60)  # Wait and retry
        result = qwed.verify(query)

Security

Is my data secure?

Yes. QWED:

  • Uses encrypted connections (HTTPS/TLS)
  • Doesn't store verification queries by default
  • SOC 2 Type II compliant (Enterprise plan)
  • Supports on-premise deployment (Enterprise Pro+)

Can QWED access my database/files?

No. QWED only sees:

  • The query you send
  • SQL schema (if verifying SQL)
  • Code snippet (if verifying code)

It cannot access your application database or files.

How long is data retained?

Default: 30 days for audit logs
Enterprise: Configurable (90 days - 7 years)
On-premise: You control retention

Performance

How fast is QWED?

Average response times:

  • Simple queries: 1-2 seconds
  • Complex queries: 2-5 seconds
  • Batch processing: 0.5s per item

Factors affecting speed:

  • Query complexity
  • Network latency
  • Verification engine used

Can I make QWED faster?

Yes:

  1. Use batch processing:

    results = qwed.verify_batch(items)  # Faster than individual calls

  2. Cache results:

    @cache
def cached_verify(query):
    return qwed.verify(query)

  3. Use async:

    results = await qwed.verify_async(queries)

Verification domains

What can QWED verify?

Supported domains:

  1. Math - Calculations, equations, algebra
  2. Logic - Propositional logic, SAT/UNSAT
  3. Code - Security vulnerabilities, syntax
  4. SQL - Injection attacks, tautologies
  5. Facts - Multi-source consensus
  6. Stats - Statistical claims
  7. Images - Visual verification
  8. Consensus - Multi-model agreement

What can't QWED verify?

Not supported:

  • Creative writing quality
  • Subjective opinions
  • Future predictions
  • Unstructured text summaries

Use cases: QWED is for objective, verifiable claims only.

Errors and debugging

Why is my verification failing?

Common causes:

  1. Malformed input:

    # ❌ Too vague
result = qwed.verify("calculate something")

# ✅ Specific
result = qwed.verify("Calculate 15% of 200")

  2. Wrong verification method:

    # ❌ Wrong
result = qwed.verify(code_snippet)

# ✅ Correct
result = qwed.verify_code(code_snippet, language="python")

  3. Network issues:

    result = qwed.verify(query, timeout=60)  # Increase timeout

How do I debug verification failures?

Enable verbose mode:

client = QWEDClient(api_key="...", verbose=True)
result = client.verify("2+2=4")
# Prints internal flow to console

Check trace:

result = client.verify("2+2=4", return_trace=True)
print(result.trace)
# Shows LLM extraction → verification steps

Production deployment

Is QWED production-ready?

Yes. QWED is used in production by:

  • Financial institutions (loan calculations)
  • Healthcare AI (drug interaction checking)
  • Legal tech (contract analysis)
  • EdTech (student assessment)

How do I deploy QWED to production?

See Production deployment guide for full checklist.

Quick steps:

  1. Test thoroughly in staging
  2. Deploy with feature flag
  3. Start with 5% traffic (canary)
  4. Monitor metrics
  5. Gradually increase to 100%

What if QWED goes down?

Recommended:

  1. Implement fallback mechanism
  2. Cache recent results
  3. Graceful degradation
try:
    result = qwed.verify(query, timeout=5)
except QWEDError:
    # Fallback to cached/approximate result
    logger.warning("QWED unavailable, using fallback")
    return fallback_result()

Support

How do I get help?

Community (free):

Enterprise support:

  • 📧 Email: support@qwedai.com
  • 💼 Slack Connect (Enterprise customers)
  • 📞 Emergency Hotline (Enterprise Pro+)

Response times:

  • Community: Best effort
  • Pro: 24-48 hours
  • Enterprise: 4-hour SLA
  • Enterprise Pro+: 1-hour SLA

Can I request new features?

Yes. Submit feature requests:

Most requested features:

  • Real-time streaming verification (Q2 2026)
  • Client-side verification (Q3 2026)
  • More language SDKs (ongoing)

Still have questions?