LLM apps

I have created multiple utility applications on Streamlit using LLMs with the help of Langchain framework. A brief of each of the service along with the components used to build it are described below.
Langsmith has been used to track all the interactions with the app.

Question-Answering

This application provides an interface for the user to ask a question and get back a response. An option is provided to choose the method of selecting the preferred LLM. The current options are:

1. OpenAI: gpt-3.5-turbo model by calling OpenAI API
2. Ollama: llama3-8b-instruct model running on the local system through Ollama
3. Groq: gemma-7b-it model by calling the Groq API
4. HuggingFace: Mistral-7B-Instruct-v0.2 model by calling HuggingFace API.

Prompt templates, output parsers and chains from Langchain are used to create an end-to-end question-answering pipeline.

Website Search

This application provides an interface for the user to essentially ask a question to a website and get a response. An option is provided to select which website to chat with. Currently, these websites are supported.

1. Wikipedia
2. Langsmith
3. Arxiv

For these websites, by default, the model is selected as OpenAI (gpt-3.5-turbo).
To develop a specialization for the model for a particular website, specialized Langchain Tools and Agents have been created.

When given a query, the agent uses the provided tool to look up for the information related to the query and generates the context without any human intervention. Based on the generated context, the agent uses the power of LLM and returns a response.
This is the power of AI Agents. They do not require any human intervention, and can perform all the steps on their own if defined properly, and if the problem is within the scope of the tools.

RAG App

This application provides capability for the user to upload multiple documents and chat with them.
It is developed by chaining PyPDFLoader, RecursiveCharacterTextSplitter, Chroma (vector database) using Retrievers

When the documents are uploaded, their content is read by the PyPDFLoader, which is then splitted into chunks by RecursiveCharacterTextSplitter. The split documents (or chunks) are passed through an Embedding model to be stored into a vector database (Chroma). When the user enter a query/question, the app performs a similarity search across all the chunks and returns the most similar ones as a context for the LLM. The LLM then responds to the question based on the context.

---

This project is still in development. More functionalities will be added in the future.