LLMR 0.6.4

New Features

Added soft structured tag output via llm_mutate_tags() and .tags in llm_mutate().
Added opt-in Gemini Vertex AI support through provider = "gemini", vertex = TRUE.

Fixed live tests to use public llm_config(api_key = "OPENAI_API_KEY") behavior.
Removed an unused internal request-body assignment that triggered a lint warning.
Updated examples to use current, inexpensive model identifiers and avoid custom temperature where models may reject it.

Excluded local Cursor files and built vignette HTML from source builds.
Improved package help, README, and vignettes for llm_mutate() shorthand, tag extraction, robust calls, and helper selection.

ollama is now supported as a provider, both for generative and embedding models. This, of course, requires a local ollama server to be up.

get_batched_embeddings was giving ugly column names. Now they go from v1 to vn

setup_llm_parallel() changed parameter order to accept numeric positional argument for workers.
llm_mutate() adds a shorthand form: llm_mutate(answer = "<prompt>" | c(system=..., user=...), .config=...).
llm_mutate() gains .structured flag: set .structured = TRUE to enable JSON output with automatic parsing (equivalent to calling llm_mutate_structured()).
llm_mutate_structured() now supports shorthand syntax: llm_mutate_structured(result = "{text}", .schema = schema).

Enhanced .fields documentation to clarify auto-extraction behavior and nested path support.
Clarified that .schema = NULL enables JSON mode without strict schema validation.
Added comprehensive examples demonstrating new structured output features in vignettes.

Removed unused internal function .category_from_condition in parallel utilities.
Fixed build_factorial_experiments() documentation to correctly describe return value columns.
Corrected call_llm_par() default value for backoff_factor parameter (now correctly documented as 3).
Added missing @importFrom purrr declarations for imported functions.

Returns and objects:
call_llm() in generative mode now returns an llmr_response object by default. Use as.character(x) to extract text; print(x) shows a concise status line; helpers include finish_reason(), tokens(), and is_truncated().
Legacy json= arguments are removed. Generative calls always return an llmr_response.

Improved API key handling.
1. You can give the key name instead of actual token
2. Even if actual token is provided, it is immediately turned into a system variable and not stored directly
json output and fixed structure are now possible.
llm_mutate can inject multiple pieces of text from columns.