features

Features engineering for structured generation for formal protocols.

mission

An operational framework for controlled generation to keep developing the standard, writing validation tests, testing hypothesis.

The main idea is to provide a set of adaptors for all the calls that are bound outside the boundary of a system towards the system that has to be make interoperable.

this repo

Create a workbench for features engineering to model inputs and outputs from publicly available LLMs.

Running the examples

This repo uses openly available models running locally using ollama. Same workflows can apply to any provider in a relatively straighforward way thanks to the langchain API.

install ollama

$ curl -fsSL https://ollama.com/install.sh | sh

Check localhost:11434 in your browser.

models selection

$ ollama pull <MODEL>
$ ollama serve
$ ollama run <MODEL>

Models available: link.

install package

$ conda create -n features python=3.12
$ conda activate features
$ cd features_bench && pip install -e .[dev]
$ git clone https://gitlab.com/masr/bspl.git
$ pip install -e ./bspl

run visualisation

It is possible to run a visualisation of the results using streamlit:

$ streamlit run features_bench/app/__init__.py

Objectives

To demonstrate a process for the definition of (standard-assuring) Web agents programmed for different tasks required to achieve the mission statement of the group. Main points tackled:

1. Layout for directories and files involved in context architecture
2. Content, format and language for the files defining the context for the "specialised agents" (agents tasked with carrying on the operations needed for the interoperability to be possbile, "interop agents" or any appropriate naming to be defined)
3. Code (for now LangChain) and workflows (for now LangGraph and LlamaIndex) for the "specialised agents". Miminal Viable Implementation to nudge best practices and provide a test bed for "specialised agents"
4. Evaluation and Testing

Working hypothesis

General assumptions:

• use interaction-oriented programming (BSPL)
• define input payloads that allow repeatable experiments
• READMEs are for human operators (so to not mingle documentation with inputs)