Releases · google/langfun

Major release for the accumulated changes since 2023/09.

Key features:

An object-in-object-out interface for LLMs with lf.query.
- more extensions: lf.complete, lf.parse, lf.describe.
LLMs under a single interface
- support various models Gemini, GPT, Claudi, Groq, and etc.
- support prompt/response caching
- support rate-limit control and automatic backoff
- support usage tracking with lf.track_usages context manager
Flexible prompt formatting with lf.Template.
- jinja2 syntax
- support drop in of multi-modality objects: image, video, PDF, and etc.
- support extension through both composition and inheritance.
Batch processing with lf.concurrent_map.
- support both order and non-ordered output.
- optional progress reporting support for console/Jupyter Notebook
Toolkit for code generation through lf.coding
sandbox for code execution via lf.coding.run
lf.PythonCode for structured code generation
FunEval: Langfun's Evaluation framework (lf.eval)
- supports dimension combination with PyGlove's pg.oneof
- supports HTML summary for both metrics and individual cases.
- model-aware parallel execution across multiple subsets.