v0.72.0

PromptingTools v0.72.0

Diff since v0.71.0

Breaking Changes

Added

• Added new Gemini 2.0 Pro and Flash models to the model registry (aliases gem20p and gem20f, respectively). Added Gemini 2.0 Flash Lite Preview model (alias gem20fl) and Gemini 2.0 Flash Thinking Experimental model (alias gem20ft).
• Added BM25 normalization kwarg to RAGTools.jl to enable 0-1 query-specific normalization of BM25 scores for easier filtering and comparison. See ?RT.bm25 and ?RT.max_bm25_score for more information.

Commits

Merged pull requests:

• Add new Google Gemini models (#266) (@svilupp)
• Max BM25 score / BM25 normalization (#267) (@svilupp)
• Update NEWS log (#268) (@svilupp)

v0.71.0

PromptingTools v0.71.0

Diff since v0.70.0

Breaking changes

Added

• Added the new OpenAI's O3 Mini reasoning model to the model registry (alias o3m).
• Added the new DeepSeek R1 Distill Llama 70b model hosted on GroqCloud, which beats GPT 4o in many benchmarks while being upto 10x cheaper (alias glmr - stands for Groq Llama Medium(70b) Reasoning).
• Added experimental support for "thinking tokens" that can be found with DeepSeek API and the reasoning model R1 (alias dsr). If the thought chain is provided, we save it in the msg.extras[:reasoning_content] field for advanced users.

Commits

Merged pull requests:

• Add o3 mini (#264) (@svilupp)
• RAGTools.jl migration warning (#265) (@svilupp)

v0.70.0

PromptingTools v0.70.0

Diff since v0.69.1

Breaking Changes

Added

• Added new DeepSeek V3 model via Deepseek (alias dschat or simply ds, because they are in a category of their own), Fireworks.ai and Together.ai (fds and tds for hosted Deepseek V3, respectively). Added Qwen 2.5 Coder 32B (alias fqwen25c or tqwen25c for Fireworks.ai and Together.ai, respectively).
• Added the reasoning Qwen QwQ 32B hosted on Together.ai.
• Added the new OpenAI's O1 model to the model registry (alias o1).
• Added DeepSeek Reasoner model alias (dsreason or simply dsr) from DeepSeek
• Added MiniMax model support and alias (minimax) for MiniMax-Text-01 model

Commits

Merged pull requests:

• Add new models (#260) (@svilupp)
• Adding Deepseek R1 model (#262) (@Sixzero)
• Add MiniMax model support (#263) (@Sixzero)

v0.69.1

PromptingTools v0.69.1

Diff since v0.69.0

Fixed

• Added assertion in response_to_message for missing :tool_calls key in the response message. It's model failure but it wasn't obvious from the original error.
• Fixed error for usage information in CamelCase from OpenAI servers (Gemini proxy now sends it in CamelCase).

Commits

Merged pull requests:

• Add assertion for aiextract (#257) (@svilupp)
• Fix usage keys for GoogleAI (#258) (@svilupp)

Closed issues:

• Update Gemini Experimental model to 1206 (#253)

v0.69.0

PromptingTools v0.69.0

Diff since v0.68.0

Added

• Added a new Gemini 2.0 Flash Experimental model (gemini-2.0-flash-exp) and updated the alias gem20f with it.

Commits

Merged pull requests:

• Add Gemini 2.0 Flash experimental (#256) (@svilupp)

v0.68.0

PromptingTools v0.68.0

Diff since v0.67.0

Added

• Added a new cache=:all_but_last cache strategy for Anthropic models to enable caching of the entire conversation except for the last user message (useful for longer conversations that you want to re-use, but not continue). See the docstrings for more information on which cache strategy to use. Thank you to @Sixzero for this!

Commits

Merged pull requests:

• Feature :all_but_last cache (#254) (@Sixzero)
• Tag new version v0.68 (#255) (@svilupp)

v0.67.0

PromptingTools v0.67.0

Diff since v0.66.0

Added

• Added a new Gemini Experimental model from December 2024 (gemini-exp-1206) and updated the gemexp alias to point to it.

Commits

Merged pull requests:

• Add Gemini experimental model 1206 and update alias (#252) (@devin-ai-integration[bot])

v0.66.0

PromptingTools v0.66.0

Diff since v0.65.1

Added

• Added support for Groq's new Llama3.3 models. Updated gllama370, gl70, glm aliases to llama-3.3-70b-versatile and added gl70s, glms aliases to llama-3.3-70b-specdec (faster with speculative decoding).

Fixed

• Fixed a bug in extract_docstring where it would not correctly block "empty" docstrings on Julia 1.11.

Commits

Merged pull requests:

• update CI for 1.11 (#249) (@svilupp)
• Add Llama 3.3 (#250) (@svilupp)

v0.65.1

PromptingTools v0.65.1

Diff since v0.65.0

Fixed

• Removed unnecessary printing to stdout during precompilation in precompile.jl.
• Fixed a "bug-waiting-to-happen" in tool use. to_json_type now enforces users to provide concrete types, because abstract types can lead to errors during JSON3 deserialization.
• Flowed through a bug fix in StreamCallback where the usage information was being included in the response even when usage=nothing. Lower bound of StreamCallbacks was bumped to 0.5.1.

Commits

Merged pull requests:

• Fix to-json-type (#248) (@svilupp)

Closed issues:

• TTFX optimization of aigenerate("whats the meaning of life"; model) (#236)
• StreamCallback issue if :usage is nothing in PT. (#247)

v0.65.0

PromptingTools v0.65.0

Diff since v0.64.0

Breaking

• Changed the official ENV variable for MistralAI API from MISTRALAI_API_KEY to MISTRAL_API_KEY to be compatible with the Mistral docs.

Added

• Added a new Gemini Experimental model from November 2024 (gemini-exp-1121 with alias gemexp).
• Added a new AnnotationMessage type for keeping human-only information in the message changes. See ?annotate! on how to use it.
• Added a new ConversationMemory type (exported) to enable long multi-turn conversations with a truncated memory of the conversation history. Truncation works in "batches" to not prevent caching. See ?ConversationMemory and get_last for more information.

Updated

• Changed the ENV variable for MistralAI API from MISTRALAI_API_KEY to MISTRAL_API_KEY to be compatible with the Mistral docs.

Commits

Merged pull requests:

• [Breaking] Renamed MISTRALAI to MISTRAL everywhere (#231) (@Sixzero)
• update Google Gemini model (#237) (@svilupp)
• feat: Add ConversationMemory and enhance AnnotationMessage (#238) (@devin-ai-integration[bot])
• Export ConversationMemory (#246) (@svilupp)