OgbujiPT 0.9 release musings

Large Language Models (LLMs) are trained with general text from around the web, but we often need them to focus on a specific real world body of facts, to help reduce hallucinations, and other issues. Retrieval Augmented Generation (RAG) is a common technique for providing this body of facts through a database, and OgbujiPT includes tools to support RAG workflows.

This has been a major area of improvement in the newly released OgbujiPT 0.9.0 0.9.1¹. Breaking down large text docuemnts into chunks for later retrieval is a key technique in RAG, and the support libraries for this have been improved. We’ve also simplified while generalizing helper classes to store content in the PGVector extension to the PostgreSQL DBMS.

One of the sound practices we try to encourage through OgbujiPT is to separate the specification of language used in LLM prompts from the processing code for those models. Word Loom is an open spec Uche Ogbuji created for managing language for AI prompting, and OgbujiPT includes a helper class, wordloom.language_item for incorporating language from Word Loom files. This is another area of significant enahncements and fixes.

The core principles of Word Loom are:

• Separation of code from natural language (straightforward process to translate any natural language elements)
• Composability of natural language elements
• Friendliness to mechanical comparisons (i.e. via diff)
• Friendliness to traditional globalization (G11N) techniques

OgbujiPT is a client-side toolkit for using large language models (LLMs), which we use every day in our consulting and products. You can use it with private, self-hosted models, or with the likes of ChatGPT, Claude and Bard. The latest release, 0.9.1 is available on GitHub and The Python Package Index (PyPI). See the changelog for further detail.

1

We quickly released version 0.9.1 to fix a few issues with demos and a hiccup in the .whl code distribution package.