Langchain-Projects

There are two projects developed by LangChain in this repo

1. Pinecone Documentation Bot
2. Python Code Interpreter
3. Advanced RAG using LangGraph(corrective RAG and self RAG) implementation

Pinecone Documentation Bot through LangChain

folder: documentation-bot-helper

The goal of this project is to utilize RAG technique to have a private bot, utilizing the LLM to answer the relevant pinecone information which we have fed to the pinecone vector database beforehand.

The specifications are: Embeddings: OpenAI text-embedding-3-small LLM model: gpt-3.5-turbo Vector Database: Pinecone Client-Server: Streamlit Memory: Chat History (Buffer Memory) Web Crawl: Firecrawl service (to convert the pinecone documentational web page to markdown and feed to the pinecone database)

The disadvantage of using buffer memory is that as conversations grow longer, the token limit of the LLM can be exceeded, truncating older parts of the chat.

input this command to get the bot running

streamlit run main.py

Result: Fig. A bot has been specialized to answer the pinecore relevant information.

Python Code Interpreter (LangChain)

Folder: slim-code-interpreter

The goal of this project is to utilize the capabilities of the ReAct agent and functional calling based on the LangChain with OpenAI model.

• Python agent: Generate the QR codes with corresponding URL and save in the hard drive.

• CSV agent: Analyze a csv data, given user prompts, e.g. find the number of books published for each author, and sort in the ascending order.

• Functional calling of Tavily search api.

A single composite agent is capable of selecting between the tools using a high-level description or logic. Use LangChain’s AgentExecutor to integrate the tools and provide an interface for the unified system. The agent prompt is taken from langchain-ai/react-agent-template

Result: Fig. After two consecutive Tavily api searches, the current weather is shown as illustrated in LangSmith

Fig. The inference process shown in console to analyze the csv file

Fig. The inference process shown in console to generate QR code and save in the disk.

Advanced RAG using LangGraph with corrective and self RAG technique

Fig. The graph representation of the flow

This project is going to help us get more quality answers when performing retrieval.

We first want to start by taking our query, performing in the vector search, semantic search, and retrieving relevant documents from our vector store. After we have all those documents, we want to start in, self-reflect, to critique those documents, and to determine whether they are indeed relevant to our original query or not.

If all of the documents are relevant to our query, then this is a happy flow and pass to LLM to generate the answer. However, documents that we find out that are not relevant to our query, we will filter them out, and perform some external web searches to obtain more information. This augment our prompt with real time information. After that, we can further verify whether an LLM generation is grounded in or supported by a sets of retrieved facts (to see if LLM was hallucinating).