Changes:

• Logging fixes/improvements
• Autofixing dialogue structure for Anthropic API

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.10.1...v3.10.2

default logging refactoring / improvements:

• possibility to configure logging output method / request+response formatters separately
• fix color reset bug

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.10.0...v3.10.1

What's Changed

• New feature Tiktoken usage for estimating number of tokens in prompt / response, fitting semantic search results to target token number
  Examples:

  // limit semantic search results
  mc.texts.search("test_collection", "query", n_results=20).fit_to_token_size(max_tokens=5, min_documents=3)
  // extended string now have .to_tokens() and .num_tokens() methods:
  prompt= mc.tpl('my_template.j2')
  tokens = prompt.to_tokens(for_model="gpt-4") 
  prompt.num_tokens()
  mc.llm(prompt).num_tokens()
  

  For usage with any regular strings, see microcore.tokenizing

• ui improvements:

  • ui.ask_yn() now have "default" argument and also handles KeyboardInterrupt
  • new function: ui.warninig()

• Other minor fixes

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.9.1...v3.10.0

bugfix: avoid recursion when json.JSONEncoder fallbacks to default & produce exception, provide original exception

Changelog

• Config.display(): Improved masking of private keys.
• Config.display(): Resolved issue #18 where the google_vertex_response_validation field appeared in config.describe for non-Google APIs.
• Config.display(): Fixed an issue where fields loaded into ENV that differ from defaults were not displayed.
• get_bool_from_env(): Corrected to ensure default values are not converted to boolean.
• Added a metrics context manager.
• Python code execution: Added max_execution_time to the Timeout error text in python.execute() to display maximum execution time.
• Updated and extended requirements, including chroma_db and anthropic.
• OpenAI API: Enhanced error handling.
• Transformers: Added support for OpenAI-style parameters (n, seed, stop).
• Fix #22: Corrected an issue where responses from the Anthropic API had a "content" property that overrides the default one provided by microcore in the LLMResponse class.
• configure(): Updated to accept a configuration file name as its sole argument.
• Enhanced capability to configure properties with list or dict types from JSON-serialized values in environment variables or configuration files.
• Enabled the configuration of callables as strings using the format <module_name>.<func_name>.
• The following configuration fields can now be loaded from environment variables or configuration files: CHAT_MODE, GOOGLE_VERTEX_GCLOUD_AUTH, GOOGLE_GEMINI_SAFETY_SETTINGS, LLM_DEFAULT_ARGS, INFERENCE_FUNC, STORAGE_DEFAULT_FILE_EXT, EMBEDDING_DB_FOLDER, MAX_CONCURRENT_TASKS.
• Additional documentation comments have been added.
• Additional tests have been added.

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.8.0...v3.9.0

• Python code execution utils added

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.7.0...v3.8.0

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.6.1...v3.7.0

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.6.0...v3.6.1

v3.6.0: PromptWrapper now additionally stores its arguments as properties

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.5.0...v3.6.0

Parallel inference functionality added

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.4.0...v3.5.0

Changes include transformers adapter improvements, new config options: & utils:

• support inference via transformers.pipeline
• init_params.always_clear_mem: gc.collect() & torch.cuda.empty_cache() before inference
• init_params.gradient_checkpointing: bool
• init_params.pipeline_task: str
• init_params.use_pipeline: bool
• possibility to configure transformers inference function by name (str)
• utils: ui.black() added
• bugfix: prevent config validation errors for ApiType.NONE
• config validation tests, Qwen1.5-1.8B-Chat added to local transformers tests

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.3.1...v3.4.0

• utils.dedent() implemented, automatically removing indent from template strings in mc.prompt()
• fmt(): alias to prompt()
• fmt, prompt, Config added mo module exports

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.2.1...v3.3.1

Changes

• ApiType.NONE added for handling setups without language model

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.2.0...v3.2.1

Changes:

• New functionality: utils.extract_number() // LLMResponse.parse_number(): parsing numbers from llm responses
• LLMResponse.gen_duration: float attribute added (inference duration in seconds)
• Configuring improvements: Automatically convert configuration keys to uppercase
• Configuring improvements: Add LLM_ prefix to keys if necessary
• Configuring improvements: Allow to configure from Config instance or dictionary
• Fixed: various occurrences of inoptimal arguments merging (impossibility to override defaults)
• Local models functionality reworked

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.1.0...v3.2.0

Changelog

• Fixing invalid JSON generated by language models

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.0.1...v3.0.2

Full Changelog: https://github.com/Nayjest/ai-microcore/compare/v3.0.0...v3.0.1

Changelog

• Support of local language models through custom inference function
• Support of local transformers models
• INIT_PARAMS: dict config parameter introduced to customize local models / API clients
• CHAT_MODE: bool config parameter introduced to force switch to text-completion/chat mode
• Config instance now may be casted to dict (convenient to reconfigure / export)
• More tests added
• API tests fixed
• get_vram_usage(), show_vram_usage() & is_google_colab() utility functions implemented
• (llm response str).as_message + (prompt | llm response str).as_user, .as_system, .as_assistant, .as_model properties added (utils.ConvertableToMessage class)

Changelog

Fix #14 : Unwanted behaviour: previous configuration remains active when calling configure(...) with invalid settings

microcore v2.1.0

• is_notebook(), is_caggle(), coloring utility functions added
• bugfix colorama in Jupyter notebooks
• more informal exceptions
• code style fixes