APPL v0.2.0: Major Improvements Compared to v0.1.2

🚀 Key Updates

1. Automatic Initialization and Enhanced Configuration

• Auto Initialization: No need to call appl.init() to initialize APPL.
• Configuration Updates:
  • Use appl.yaml to overwrite the configs is still effective, with warning on items not in the config (useful for checking wrong configs).
  • Use appl.init(**kwargs) or command line for runtime updates.
  • Introduced Pydantic models for stricter type checking of configs and global_vars.
  • Command-line support via jsonargparse (see cmd args example).

2. Advanced Tracing and Visualization

• Langfuse and Lunary Integration:
  • Visualize traces with metadata such as git info and command-line arguments.
  • Store code for ppl and @traceable functions in trace for viewing in Langfuse.
  • New print_trace function to send traces, can be called at the end.
• appltrace Command: Export traces to multiple formats:
  • Langfuse (recommended), Lunary, HTML, Chrome tracing, etc.

3. New Features

• Explicit Growing Prompts:
  • Use grow to explicitly grow the prompt.
• Independent gen Calls:
  • appl.gen (or appl.completion) can be called outside of ppl function with messages, behavior like a wrapper around litellm.completion.
• Auto Continuation:
  • Automatically continue the generation stopped due to length limit. Ask for repeating the last line and concatenate by overlapping to achieve a reliable continuation.
• LLM Call Caching:
  • Add persistent Database for caching LLM calls with temperature=0.
• Concurrency:
  • Add global executor pool with configurable max_workers to limit the number of parallel LLM calls.
• Better Image and Audio Support:
  • Support pillow's Image as part of prompt.
  • Support Audio as part of the prompt.

4. More Examples

• Tree-of-Thought:
  • Add example for tree-of-thought with much less codes and enabled parallelization with about 6x speedup.
• Virtual Tool:
  • Add example for using LLM to emulate tool execution.
• Streamlit:
  • Add example for using streamlit to build a chat web-app with APPL.
• Command Args:
  • Add example for using command args to update configs.
• Cursor Usage [Experimental]:
  • Add example cursor rules for better using cursor to generate APPL codes.

5. Updated Default Configs

• Defaults:
  • Default server is now None (explicit setup required via appl.yaml, command line, or appl.init(servers=...)). You can use model name as server name without setting the server configs.
  • Logging: Enabled file logging by default but disabled logging LLM call arguments by default.
  • Streaming: default to print the streaming output, was live using rich.Live before.

6. Miscellaneous Fixes

• Minor bugs fixed and improved code readability.
• Code refactored for better maintainability.

Full Changelog: v0.1.2...v0.2.0