Make temperature configurable per extraction

[michaeldistel]

Problem

Temperature is currently hardcoded to 0.0 in LLMClient:

this.config = {
    temperature: 0.0,  // hardcoded!
    maxTokens: 2000,
    timeout: 120000,
    ...config,
};

While 0.0 is good for deterministic extraction, some use cases might benefit from slightly higher values for:

• Creative interpretation of ambiguous text
• Fuzzy matching
• Handling typos and variations

Proposed Solution

Allow temperature override at multiple levels:

1. Global config (already works, but overridden):

const client = new LLMClient({
  baseURL: '...',
  model: '...',
  temperature: 0.2,  // custom default
});

2. Per-extraction override:

await client.extract({
  schema,
  input,
  temperature: 0.1,  // one-off override
});

3. Schema-level default (via metadata):

{
  "metadata": {
    "temperature": 0.2
  }
}

Implementation

Fix LLMClient constructor to respect passed temperature:

this.config = {
    ...config,
    temperature: config.temperature ?? 0.0,
    maxTokens: config.maxTokens ?? 2000,
    timeout: config.timeout ?? 120000,
};

Add temperature to ExtractionOptions:

interface ExtractionOptions {
  schema: Schema;
  input: string;
  systemPrompt?: string;
  temperature?: number;  // add this
}

Acceptance Criteria

• Config-level temperature is respected
• Per-extraction temperature override works
• Tests verify configuration precedence
• Documentation explains when to adjust temperature
• Default remains 0.0 for determinism

Related

Parent: #30