mLLMCelltype 2.0.8

Reliability fixes

• Disabled DeepSeek V4 thinking mode for cell-type annotation requests in both
  the R and Python packages. This prevents the response budget from being
  exhausted before final cell-type labels are returned.
• Increased API response timeouts for slower non-streaming DeepSeek requests:
  the R default is now 120 seconds and the Python DeepSeek timeout is 120
  seconds.
• Unified R prompt preparation across plain and structured reasoning modes.

The DeepSeek fixes were validated against the current API behavior and by a
real mLLMCelltype user who confirmed that annotation now completes normally.

Validation

• R: 690 tests passed; 4 integration tests skipped by default.
• Python: 569 tests passed; 1 integration test skipped by default.
• R and Python source distributions built successfully.
• Python wheel and source distribution passed twine check.

Installation

install.packages("mLLMCelltype")

Until CRAN publishes 2.0.8, install the R package directly from GitHub:

remotes::install_github("cafferychen777/mLLMCelltype", subdir = "R", force = TRUE)

pip install --upgrade mllmcelltype

Full changelog: v2.0.7...v2.0.8

mLLMCelltype 2.0.7

mLLMCelltype 2.0.7

New Features

• Kimi (Moonshot AI) provider with dual-protocol support. Models prefixed with kimi-/moonshot- (default kimi-k2.6) resolve automatically and authenticate via MOONSHOT_API_KEY (or KIMI_API_KEY); the Moonshot Open Platform uses OpenAI-compatible Chat Completions, and Kimi Code /messages endpoints auto-select the Anthropic Messages protocol.
• return_reasoning option for annotate_cell_types() (R) / annotate_clusters() (Python), returning a structured per-cluster cell_type / marker_genes / gene_expression record instead of a plain label.

Bug Fixes

• Annotation parsing corrected so common LLM output shapes no longer produce wrong or shifted annotations on 0-based (Seurat/Scanpy) clusters: numbered lists (1., 2., …), preamble/header lines, annotation-internal colons, a mid-list Unknown, a stray Summary:-style line, and out-of-order explicit labels.
• MiniMax transient in-body errors (rate limits / timeouts reported as HTTP 200 + non-zero base_resp.status_code) are now retried instead of silently dropping the model from the consensus panel.
• Anthropic responses that lead with a non-text (thinking) block no longer discard the answer — every text block is concatenated and non-text blocks skipped.
• is_error_response no longer discards a valid multi-cluster response over a single Error: line.
• A warning is emitted when a provider response is truncated at max_tokens.

This release also ships the previously-unreleased 2.0.6 reliability work (unified validation, centralized retries, deterministic parsing/caching, normalized usage telemetry).

The Kimi provider is verified against the Moonshot/Kimi API documentation and covered by mocked tests; a live smoke test against the real API is recommended for Kimi-specific usage.

mLLMCelltype 2.0.5

Summary

• Release R and Python package metadata for version 2.0.5.
• Include discussion parsing, Qwen endpoint cache scoping, and custom provider configuration fixes from the latest main branch.
• Publish Python 2.0.5 artifacts and attach the R source package prepared for CRAN submission.

Verification

• Python: python -m ruff check .
• Python: python -m pytest — 318 passed, 1 skipped
• R: devtools::test() — 207 passed, 4 skipped
• R: R CMD check --as-cran mLLMCelltype_2.0.5.tar.gz — completed with 2 NOTEs

mLLMCelltype v2.0.4

mLLMCelltype v2.0.4

Release archived on Zenodo for citation in:

Yang, C., Zhang, X. & Chen, J. Large Language Model Consensus Substantially Improves the Cell Type Annotation Accuracy for scRNA-seq Data. Communications Biology (2026).

Changes since v2.0.0

• Hardened consensus and report paths
• Fixed report metadata model normalization
• Aligned direct discussion flow with consensus-model key autoload
• Improved missing-value handling and discussion seed sanitization

mLLMCelltype v2.0.0

mLLMCelltype v2.0.0

Breaking Changes

• Logger defaults changed: UnifiedLogger console output is now disabled by default. Use configure_logger(console_output = TRUE) to enable.

New Model Support

• OpenAI: GPT-5.2, GPT-5.1, GPT-5, o3-pro, o4-mini
• Anthropic: Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5
• Google: Gemini 3 Pro, Gemini 3 Flash, Gemini 2.5 Pro/Flash
• X.AI: Grok 4, Grok 4.1, Grok 4 Heavy
• DeepSeek: DeepSeek R1
• Alibaba: Qwen 3 Max
• Zhipu: GLM-4.7
• MiniMax: MiniMax M2.1

Architecture Improvements

• Unified logging: consolidated dual debug mechanisms into single log_debug() approach
• Complete audit trail: all warning() calls paired with log_warn(); LLM consensus decisions logged
• NAMESPACE fix: QwenProcessor correctly exported; .qwen_endpoint_cache removed from public API
• Cluster ID preservation throughout annotation pipeline

Bug Fixes

• Fixed @export misplacement exporting internal state instead of QwenProcessor class
• Fixed missing tissue_name parameter in API integration tests
• Fixed print() bypassing logging system in model comparison
• Fixed trailing newlines causing double-spaced console output
• Improved NA handling in API key validation

Package Info

• R Package: mLLMCelltype_2.0.0.tar.gz
• Tests: 174 PASS, 0 FAIL, 0 WARN

mLLMCelltype v1.2.9: Cache System Fix and Improvements

mLLMCelltype v1.2.9 Release Notes

🎉 Major Updates

Cache System Fix (#65)

• Fixed critical cache isolation issue: OpenRouter models were incorrectly reusing cache from regular models
• Solution: Modified create_cache_key() function to properly normalize provider for OpenRouter models (models containing '/')
• Impact: Prevents cache pollution and ensures accurate cell type annotations when switching between providers

New Features

• Cache Management Module: Added cache_manager.py with utilities for cache inspection and management
  • get_cache_info(): Get information about current cache state
  • clear_mllmcelltype_cache(): Interactive cache clearing
  • CLI interface: python -m mllmcelltype.cache_manager
• Comprehensive Documentation: Added detailed cache system documentation in docs/CACHE_SYSTEM.md
• Example Scripts: Added cache management examples in examples/cache_management_example.py

Code Quality Improvements

• Fixed all ruff linting errors across the Python package
• Added .gitignore to exclude temporary test files
• Updated import ordering and removed unused imports
• Added comprehensive test suite for cache system validation

📦 Package Information

Python Package

• Version: 1.2.9
• Files:
  • mllmcelltype-1.2.9.tar.gz - Source distribution
  • mllmcelltype-1.2.9-py3-none-any.whl - Wheel distribution

R Package

• Version: 1.2.9
• File: mLLMCelltype_1.2.9.tar.gz

🐛 Bug Fixes

• Cache isolation between regular and OpenRouter models
• Import order issues in examples
• Unused variable warnings
• F-string formatting issues

📚 Documentation Updates

• Added comprehensive cache system documentation
• Updated examples to demonstrate proper model usage
• Added API key configuration example (.env.example)

🧪 Testing

• Added extensive test suite for cache system
• Verified cache isolation with real API calls
• Tested edge cases including tissue context variations

Installation

Python

pip install mllmcelltype==1.2.9

R

install.packages("mLLMCelltype_1.2.9.tar.gz", repos = NULL, type = "source")

Acknowledgments

Thanks to @eason-analytics for reporting the cache issue (#65) and providing detailed reproduction steps!

mLLMCelltype v1.2.8

Release of mLLMCelltype v1.2.8 (R package) and v1.2.3 (Python package)

v1.2.4 - Critical Bug Fix Release

Critical Bug Fixes

Fixed Major as.logical(from) Error

• Resolved critical error that occurred when processing large numbers of clusters (60+ clusters)
• This error was caused by non-character data being passed to strsplit() functions
• Users can now successfully process datasets with many clusters without encountering type coercion errors

Enhanced API Response Handling

• Added comprehensive tryCatch() blocks around all strsplit() operations in API processing functions
• Improved response validation to prevent function/closure types from being processed as character strings
• Enhanced error handling for all API processing functions:
  • process_openrouter.R
  • process_anthropic.R
  • process_openai.R
  • process_deepseek.R
  • process_qwen.R
  • process_stepfun.R
  • process_minimax.R
  • process_zhipu.R
  • process_gemini.R
  • process_grok.R

Improvements

• Better NULL value handling: Improved unlist() operations to filter out NULL values and handle errors gracefully
• Enhanced logging: Added more detailed error logging for debugging API response issues
• Improved consensus checking: Enhanced check_consensus.R to handle edge cases with malformed responses

Technical Details

• Fixed issue where large cluster datasets could cause type coercion errors in response parsing
• Added validation for function/closure types in API responses to prevent downstream errors
• Improved error messages to provide better diagnostics for API response issues

Installation

# Install from GitHub
devtools::install_github("cafferychen777/mLLMCelltype", subdir = "R")

What's Changed

• Enhanced robustness when processing large datasets
• Improved error handling across all API integrations
• Better debugging capabilities with enhanced logging

Full Changelog: v1.2.3...v1.2.4

mLLMCelltype 1.2.0

1.2.0 (2025-04-28)

Improvements

• First CRAN release
• Updated documentation and vignettes
• Improved package structure for CRAN compliance
• Fixed CITATION file format
• Updated URL to https://cafferyang.com/mLLMCelltype/

v1.1.4 Release

mLLMCelltype v1.1.4 (2025-04-24)

New Features & Improvements

• Expanded OpenRouter models list to support free models
• Resolved inconsistency issues between international and mainland China versions of Qwen models

Bug Fixes

• Fixed OpenRouter model handling in the consensus check process
• Fixed undefined current_clusters variable in README files by adding explicit variable definition and comments

Other Updates

• Updated DESCRIPTION file

This version improves model compatibility and fixes several key issues. We recommend all users to update to this version.