Changelog • AIGENIE

AIGENIE 2.1.2

CRAN release: 2026-09-08

Added item.attribute.definitions to AIGENIE() and local_AIGENIE(): an optional named list of definitions for some or all attributes, inserted into the generation prompt immediately after each item type’s attribute list.

temperature and top.p now default to NULL in AIGENIE(), local_AIGENIE(), chat(), and local_chat() because many recent models reject these parameters. When NULL, they are not sent to the model, so the defaults apply.
Exposed boot.iter and ncores in AIGENIE(), GENIE(), local_AIGENIE(), and local_GENIE(). The boot.iter default remains 500, matching the current reduction pipeline; ncores = NULL preserves EGAnet’s existing default core behavior.
Added a publication-ready filtering_audit with item-level filtering provenance for UVA and bootEGA decisions.
Added per-type reduction_summary outputs.
Added pre-reduction network-loading diagnostics for filtered items.
Added bundled GPT-5.4 items and frozen embeddings for reproducible GENIE examples.
Added an official filtering-audit vignette and pkgdown documentation site.

AIGENIE() and local_AIGENIE() now honor a user-supplied EGA.model; previously it was silently ignored and both models were always compared.
With all.together = TRUE, final_items (and initial_items when keep.org = TRUE) now retain the EGA_com community column alongside the original type and attribute labels.
AIGENIE() now returns the type-level results when the run.overall fit fails, even when silently = TRUE.
AIGENIE() and local_AIGENIE() no longer error when item generation returns zero items.

Corrected stale function documentation, including the structure of the returned objects and the EGA.algorithm default ("walktrap").
Internal helpers are no longer listed in the package reference index.

Restored correct UVA redundancy detection with current EGAnet versions.
Set bootEGA reduction to 500 bootstrap iterations.
Full embeddings now win exact full/sparse NMI ties within an EGA model.
run.overall = TRUE now performs a pooled post-reduction fit without applying a second item-reduction pass.

Updated Authors@R so Lara Russell-Lasalandra and Hudson Golino are explicitly represented as package authors/co-creators and copyright holders; Alexander Christensen remains a full author and copyright holder.
Added a formal AIGENIE software citation with the CRAN package DOI.
Updated the methodology citation to the published 2026 Behavior Research Methods article.

Anthropic Claude support: Generate items using Claude Sonnet 4.5, Opus 4, and Haiku 4.5 models via anthropic.API.
Jina AI embeddings: Compute embeddings using Jina models (v2, v3, v4) via jina.API, including Matryoshka truncation and task-specific adapters.
Slash-style Groq models: Route HuggingFace-style model IDs (e.g., "meta-llama/llama-4-scout-17b-16e-instruct", "qwen/qwen3-32b") to Groq when groq.API is provided.
chat() function: Send arbitrary prompts to any supported LLM provider with optional repetitions, temperature control, and system role customization.
list_available_models(): Query available models across all providers, with optional filtering by type ("chat" or "embedding").
Multi-provider mixing: Use one provider for item generation and another for embeddings (e.g., Anthropic items + Jina embeddings).

Comprehensive input validation across all user-facing functions.
Unified embedding dispatch via generate_embeddings() supporting OpenAI, Jina AI, HuggingFace (API and local), and sentence-transformers.
Model alias resolution for common shorthand names ("sonnet", "llama3", "deepseek", "qwen", etc.).
Improved error messages with provider-specific guidance.

Fixed provider detection for slash-style model names that were incorrectly routed to HuggingFace instead of Groq.

Complete rewrite of the pipeline architecture.
Introduced GENIE() for validation of user-provided item sets (embedding → EGA → UVA → bootEGA).
Added Groq API support for open-source LLM item generation.
Added local LLM support via llama-cpp-python (local_AIGENIE(), local_GENIE(), local_chat()).
Python environment management via UV and reticulate.
HuggingFace embedding support (API and local sentence-transformers).