Skip to main content
GitHub

Parameter Fixes

Adjust LLM parameters.

Not available in this release

No new fix is generated or deployed today: fix generation runs after a diagnosis, which is held for the beta, and fix deployment is gated off. This page describes the fix type as designed.

Parameter fixes adjust LLM configuration like temperature, max_tokens, etc.

Use Cases

  • Reduce hallucination (lower temperature)
  • Get longer responses (increase max_tokens)
  • Increase diversity (adjust top_p)
  • Control repetition (penalties)

Configuration

Parameters are nested inside a parameters dict within config:

{
  "fix_id": "fix-param-001",
  "fix_type": "parameter",
  "config": {
    "parameters": {
      "temperature": 0.3,
      "max_tokens": 4096
    }
  }
}

Supported Parameters

ParameterTypeDescription
temperaturefloatSampling randomness (0-2)
max_tokensintMaximum output tokens
top_pfloatNucleus sampling (0-1)
top_kintTop-k sampling
frequency_penaltyfloatReduce repetition (-2 to 2)
presence_penaltyfloatEncourage new topics (-2 to 2)
stopstring[]Stop sequences
timeout_msintRequest timeout

Common Fixes

For Hallucination

Lower temperature for more deterministic output:

{
  "fix_type": "parameter",
  "config": {
    "parameters": {
      "temperature": 0.1
    }
  }
}

For Truncated Output

Increase max_tokens:

{
  "fix_type": "parameter",
  "config": {
    "parameters": {
      "max_tokens": 8192
    }
  }
}

For Repetitive Output

Add penalties:

{
  "fix_type": "parameter",
  "config": {
    "parameters": {
      "frequency_penalty": 0.5,
      "presence_penalty": 0.3
    }
  }
}

For Timeouts

Increase timeout:

{
  "fix_type": "parameter",
  "config": {
    "parameters": {
      "timeout_ms": 60000
    }
  }
}

Provider Compatibility

Parameters are not renamed per provider: the runtime sets exactly the keys given under parameters, as written. Use the parameter names of the provider your calls go to.

Targeting

A fix applies by its error_code — an exact code, or a pattern ending in * that matches by prefix. There is no way to target specific models.

{
  "error_code": "OUTPUT.QUALITY.REPETITIVE"
}

Next Steps