Added
- Command line interface `botex` to initialize oTree sessions and start bots from the command line
- Added LLM API throttling to deal with LLM rate limits
- botex now explicitly checks for oTree validation checks and provides this feedback to LLM bots to trigger corrected responses
- oTree pages with multiple buttons but no next button are now handled correctly
- Option to use an already existing llama.cpp endpoint
- Parsing of botex environment file by calling `load_botex_env()` function
- Convenience functions `start_otree_server()` and `stop_otree_server()` to facilitate starting oTree server from within Python
- Function `export_otree_data()` to export oTree wide multi-app data via the oTree web frontend
- Function `normalize_otree_data()` to normalize oTree wide multi-app data to a normalized set of list-of-dicts that is convenient for downstream analysis. This function is based on some educated guesses on the oTree data format and will likely need adjustment for different oTree experiment setups.
Changed
- Communication with LLMs is now based on structured outputs, requiring the use of LLM endpoints that support this feature (making gpt-4o-2024-08-06 the new standard model).
- For those models that don't support structured outputs via LiteLLM, there is a largely untested fallback using the `instructor` package. This is experimental and will likely be removed as soon as LiteLLM supports Ollama's structured output feature.
- Prompting has been improved by streamlining language and adjusting it to the usage of structured outputs.
- The API of the main functions has been standardized and moved closer to the API of LiteLLM.
- `run_bots_on_session()` and `run_single_bot()` now return the resulting thread object(s) if `wait` is set to `False`.
- botex now logs to its own logger.
Fixed
- Rearranged llama.cpp server logging to a file to avoid pipe buffer overflows.
- Removed deadlock loop on answers that were not accepted by oTree