chatRater NEWS

chatRater 1.3.1 (2026-06-25)

This is a CRAN-maintenance release. 1.3.1 adds no new user-facing features; it fixes two issues surfaced by R CMD check on CRAN incoming and tightens packaging metadata so the package clears incoming feasibility checks.

Changes relative to 1.3.0

• CRAN incoming fix — aggregate import: rate_openai_weighted() calls stats::aggregate() for mean/SD pooling. CRAN incoming flagged this as an undefined global function. Added importFrom(stats, aggregate) to the NAMESPACE (generated via roxygen tag).
• CRAN incoming fix — rank parameter documentation: alignment() accepts a rank argument (logical, default FALSE) but it was missing from the roxygen \usage{} block, triggering a Codoc mismatches WARNING. Documented rank in man/alignment.Rd (added to \usage{} and \arguments{} with @param rank Logical; ifTRUE, compute Spearman/Sigma rank correlation alongside Pearson statistics. DefaultFALSE.).
• Citation tightening: inst/CITATION is now a single APA 7 entry pointing at the PsyArXiv preprint (doi:10.31234/osf.io/mje6w_v1). The duplicate “software manual” entry was removed — citation("chatRater") now returns one tidy reference.
• Date field: Updated Date: in DESCRIPTION to 2026-06-25 to match the build date.

Verification

R CMD check is now clean on CRAN incoming (Windows Server 2022, R-devel r90190, ucrt):

Status: OK

No new warnings, no notes, no missing documentation entries.

chatRater 1.3.0 (2026-05-16)

Major Changes

• Probability-weighted scoring (method = "weighted"): New rating method following Brysbaert et al. (2025). Instead of extracting the most probable token, the function retrieves log probabilities for all candidate rating tokens and computes a continuous score as Σ(rating × p(rating)). This yields finer-grained estimates than dominant-token decoding alone. Supported for provider = "openai" with text stimuli only.
• top_logprobs parameter (1-20): Controls how many most-likely tokens are returned at each position. Default 5; set to 20 for maximum coverage of token surface variants (e.g., “3”, ” 3”, “3.”).
• include_probs parameter: When TRUE, the returned data frame includes a probs_json column with the full probability distribution over rating values as a JSON string (e.g., {"3":0.52,"4":0.43,"2":0.05}).
• Direct API call for weighted mode: rate_openai_weighted() bypasses llmcoder to make a direct httr2 call with logprobs = TRUE, enabling fine-grained probability extraction without modifying the llmcoder dependency.
• Migrated from openai to llmcoder: Core rating for text stimuli now uses llmcoder::call_llm(), enabling multi-provider support (OpenAI, Anthropic, Ollama, LM Studio, DeepSeek, Groq, Mistral, OpenRouter, OpenAI-compatible).
• Local image file support: Images can now be passed as local file paths (base64-encoded) in addition to URLs. Vision-capable models (gpt-4o, etc.) are used automatically for image stimuli.
• Audio support: Local audio files (mp3, wav, ogg, flac, m4a, aac) are now supported; transcribed via OpenAI Whisper before rating.
• return_type parameter: Control output format: "numeric" (default, extract numbers only), "text" (return full text response), "raw" (return unprocessed API response).
• columns parameter (experimental): Select which columns appear in the returned data frame. Available columns: "stim", "rating", "iteration", "scale", "type", "provider", "model". Default NULL returns all columns.

New Examples

• Weighted rating with include_probs = TRUE
• Batch weighted rating with generate_ratings_for_all()
• Expanded @examples in generate_ratings() and generate_ratings_for_all(): 9 detailed examples covering text (string), image (local file + URL), and audio (local file) stimuli, with both numeric rating and full-text description workflows.

Bug Fixes

• Fixed curl::form_file() usage in Whisper API upload (previously used incorrect httr2 multipart syntax).
• Fixed debug parameter not passed to rate_build_content_blocks(), causing "argument is not interpretable as logical" error when processing audio or image stimuli.

Dependencies

• Imports: llmcoder (>= 1.2.0), base64enc, tools, httr2, curl, jsonlite

Citation

• Added inst/CITATION with APA 7-formatted entries for both the software (Manual) and the associated PsyArXiv preprint (doi:10.31234/osf.io/mje6w_v1). citation("chatRater") returns both references.

chatRater 1.2.0 (2025-08-18)

Changes

• Removed text analysis utilities: get_lexical_coverage(), get_word_frequency(), get_zipf_metric(), get_levenshtein_d(), get_semantic_transparency() have been removed to focus the package on its core rating functionality.
• Simplified API: Only generate_ratings() and generate_ratings_for_all() are now exported.
• Image URL support: Stimuli can now be image URLs (passed to GPT-4 Vision via the openai package).

Dependencies

• Imports: openai, tools

chatRater 1.1.0 (2025-05-01)

Initial CRAN Release

• generate_ratings() for single-stimulus rating via LLM.
• generate_ratings_for_all() for batch rating of multiple stimuli.
• Text analysis utilities:
  • get_lexical_coverage(): Calculate lexical coverage of a text.
  • get_word_frequency(): Look up word frequency information.
  • get_zipf_metric(): Calculate Zipf metric for a text.
  • get_levenshtein_d(): Calculate Levenshtein distance between strings.
  • get_semantic_transparency(): Rate semantic transparency.
• OpenAI API integration (text-only stimuli).
• Rating scale validation (format: "1-7", "0-10", etc.).