Skip to content

Latest commit

 

History

History
62 lines (38 loc) · 2.01 KB

README.md

File metadata and controls

62 lines (38 loc) · 2.01 KB
Raw

BSHR Streamlit App

This is a Streamlit app that utilizes the Brainstorm-Search-Hypothesize-Refine (BSHR) pipeline to answer user queries.

Overview

The app has two main components:

  • app.py: The Streamlit frontend that takes a query input and displays the BSHR response.

  • st_components/bshr.py: Wrapper functions to call the BSHR pipeline. Provides both the full BSHR pipeline as well as a simpler "vanilla" GPT-3 response.

The main BSHR pipeline code lives in BSHR/bshr_pipeline.py. It brings together the different components:

  • Brainstormer: Generates search queries using GPT-3 based on the main query.

  • Searcher: Searches Wikipedia and a Searx instance using the generated queries.

  • InformationProcessor: Checks relevancy of results, condenses information, and generates hypotheses.

  • HypothesisTournament: Refines hypotheses through a tournament to determine the best one.

  • Responder: Crafts a conversational response using the final hypothesis.

The pipeline uses the OpenAI API, with the option to mock requests for testing using https://openrouter.ai.

Requirements

Make sure to install requirements before running the app:

pip install -r requirements.txt

Configuration

The app expects a .streamlit/secrets.toml file with the following:

[openai]
api_key="sk-..." 
api_base= ... # "https://api.openai.com/v1" or "https://openrouter.ai/api/v1"
base_model= ... # model_name 
mock_openai= ... # false for true OpenAI, true for openrouter or other openai mock servers

[search]
searx = ... # your own searx server

You need to manually create this file and fill in your own OpenAI API key and other settings.

Running the App

With the secrets configured:

streamlit run app.py

The main query can then be entered and the BSHR pipeline will be invoked to generate a response.

Customizing

The different pipeline components in bshr_pipeline.py can be configured or swapped out as needed. New question answering strategies can be built by mixing and matching the different modules.