Notable changes to Format based on Keep a Changelog. Project follows Semantic Versioning.
llm_helper.debug_model_manager—a way to extract raw prompt & schema/tool-call info for debugging of underlying LLM behavior- docs beyond the README (
docfolder) - test cases
- demo/algebra_tutor.py
- demo/blind_obedience.py
- use of logger rather than trace boolean, throughout
- further code modularizarion and reorg
- improvements to default prompting
- more elegant handling of install from an unsupported OS
- handling of multi-trip scenarios
- notes on how to override prompting
- processing for function-calling system prompts
- server startup 😬
- demo
demo/zipcode.py - support for multiple workers & CORS headers (
--workers&--cors_origincmdline option)
- async tool definitions
toolio.respondermodule, with coherent factoring fromserver.pyllm_helper.model_managerconvenience API for direct Python loading & inferencing over modelsllm_helper.extract_contenthelper to simplify the OpenAI-style streaming completion responsestest/quick_check.pyfor quick assessment of LLMs in Toolio- Mistral model type support
- Turn off prompt caching until we figure out #12
- Have responders return actual dicts, rather than label + JSON dump
- Factor out HTTP protocol schematics to a new module
- Handle more nuances of tool-calling tokenizer setup
- Harmonize tool definition patterns across invocation styles
- More vector shape mamagement
- Legacy OpenAI-style function-calling support
trip_timeoutcommand line option fortoolio_request- Support for mixtral model type
- Model loading timing
timeoutclient param totrip_timeout
- tool/param.rename, e.g. for tool params which are Python keywords or reserved words
- API example in README
- Type coercion for tool parameters
- Ability to rename params in for tools
- Three test cases, including currency conversion
- Excessive restrictions in OpenAI API
- A couple of test cases
- Error when tool is not used
- Initial standalone release candidate