For free-text tasks: writing, rewriting, summarizing, explaining. If a
machine-processable result is needed, use 'local_structured' instead.

Args:
    model: Model name as in 'list_models', for example 'llama3.2:3b'.
    prompt: The actual instruction for the model.
    system: Optional role/behavior instruction.
    temperature: 0 means as deterministic as possible (default), higher means more creative.
    max_tokens: Optional cap on the number of generated tokens.
    context: Optional text prepended to the prompt (source, excerpt, data).
    think: Enables the model's reasoning trace. Off by default, because the
        reasoning trace otherwise consumes the answer's token budget (with a
        tight 'max_tokens' the answer then comes back empty). Only for models
        with the 'thinking' capability.

Uses the respective backend's schema enforcement (Ollama's 'format' field,
vLLM's 'response_format' with 'json_schema') and then additionally validates
the answer against the schema itself. If the output does not satisfy the
schema, it retries up to twice, passing the model the concrete violations.
Only then an error, but then with the invalid raw output, so it is visible
what went wrong.

Args:
    model: Model name, for example 'llama3.2:3b'.
    prompt: Instruction on what should be extracted from which text.
    schema: JSON schema of the desired result (object with 'type', 'properties', ...).
    system: Optional role/behavior instruction.
    think: Enables the model's reasoning trace. Off by default, because the
        reasoning trace otherwise consumes the answer's token budget. Only
        for models with the 'thinking' capability.

Useful for similarity comparisons, duplicate detection or a rough sort by
topic, without troubling a large language model.

Args:
    texts: List of texts to embed.
    model: Embedding model, with Ollama preset to 'nomic-embed-text'.