Skip to content

model-selection.md

Latest commit

 

History

History
58 lines (39 loc) · 2.51 KB

model-selection.md

File metadata and controls

58 lines (39 loc) · 2.51 KB
Raw

Model Selection Guide

VXD supports multiple LLM providers and lets you assign different models to different roles based on your cost, quality, and availability requirements.

Providers

Provider Key Models Cost
anthropic ANTHROPIC_API_KEY or Claude CLI Claude Opus/Sonnet/Haiku Paid (or subscription)
openai OPENAI_API_KEY GPT-4o, GPT-4o-mini, o3 Paid
google GOOGLE_AI_API_KEY Gemma 4 27B Free tier (1500 req/day)

Execution vs Verification Tiers

VXD roles fall into two tiers with different quality requirements:

Execution Tier — High volume, code generation focus:

  • Junior (simple stories, 1-3 complexity)
  • Intermediate (medium stories, 3-5 complexity)
  • Supervisor (periodic drift checks)

These roles benefit from fast, cheap models. Gemma 4 on Google AI's free tier is ideal.

Verification Tier — Low volume, reasoning focus:

  • TechLead (requirement decomposition, planning)
  • Senior (complex stories, code review, escalations)
  • QA (plan-vs-execution verification, quality gate)
  • Manager (failure diagnosis, corrective actions)

These roles need strong reasoning to catch issues. Claude Sonnet/Opus recommended.

Default Configuration

VXD ships with a hybrid setup that maximizes free-tier usage:

Role Provider Model Tier
TechLead anthropic claude-opus-4 Verification
Senior anthropic claude-sonnet-4 Verification
Intermediate google gemma-4-27b-it Execution
Junior google gemma-4-27b-it Execution
QA anthropic claude-sonnet-4 Verification
Supervisor google gemma-4-27b-it Execution
Manager anthropic claude-sonnet-4 Verification

Choosing Models

Budget-conscious: Use the defaults. Execution roles are free, verification roles use your Anthropic subscription or API key.

Maximum quality: Put all roles on Claude Sonnet or Opus. Higher cost but strongest reasoning across the board.

Fully free: Put all roles on Google/Gemma 4. Works for smaller projects but QA and Manager may miss subtle issues.

OpenAI alternative: Replace Anthropic roles with OpenAI equivalents (GPT-4o for verification, GPT-4o-mini for execution).

Fallback Behavior

When a Google-configured role hits free-tier quota limits, VXD automatically falls back to Anthropic (Claude CLI or API). This is transparent — no manual intervention needed. See the Gemma 4 Guide for details on quota limits and fallback chain.