metadata

license: apache-2.0
title: Self-Reflective CRAG Application "Info Assistant"
sdk: streamlit
emoji: 🌍
colorFrom: blue
short_description: Self Reflective Multi Agent LangGraph CRAG Application
sdk_version: 1.38.0

Overview

This project demonstrates a self Reflective corrective Retrieval Augmented Generation (CRAG) application built using LangGraph. The application leverages a Gemma2 9B LLM to provide informative and relevant responses to user queries. It employs a multi-agent approach, incorporating various components for enhanced performance and user experience.

Key Features

Vector Store: Uses Chroma Vector Store to efficiently store and retrieve context from scraped webpages related to data science and programming. Prompt Guard: Ensures question safety by checking against predefined guidelines. LLM Graders: Evaluates question relevance, answer grounding, and helpfulness to maintain high-quality responses. Retrieval and Generation: Combines context retrieval from vector store and web search with LLM generation to provide comprehensive answers. Iterative Refinement: Rewrites questions and regenerates answers as needed to ensure accuracy and relevance. Customization: Offers flexibility in model selection, fine-tuning, and retrieval methods to tailor the application to specific requirements. Local Deployment: Can be deployed locally for enhanced user data privacy.

Technical Specifications

LLM: Gemma2 9B Vector Store: Chroma Embeddings: Alibaba-NLP/gte-base-en-v1.5 Workflow: LangGraph Model API: ChatGroq Web Search: Wikipedia and Google SERP

Workflow

User Query: User inputs a question. Prompt Guard: Checks if the question is safe and appropriate. Context Retrieval: Searches the vector store for relevant documents. Document Relevance: Evaluates document relevance using LLM graders. Web Search: If necessary, conducts web searches on Wikipedia and Google SERP. Answer Generation: Generates a response using the retrieved documents and LLM. Answer Evaluation: Evaluates answer grounding and helpfulness using LLM graders. Refinement: If necessary, rewrites the question or regenerates the answer.

Customization Options

Model Selection: Choose different LLM models based on specific needs (e.g., larger models for more complex tasks). Fine-Tuning: Fine-tune the LLM to match specific styles or domains. Retrieval Methods: Explore alternative vector stores or retrieval techniques.

Local Deployment

To deploy the application locally, follow these steps:

Set up environment: Install required dependencies (LangGraph, Chroma, LLM API, etc.). Prepare data: Scrape webpages and create the vector store. Configure workflow: Define the workflow and LLM graders. Run application: Execute the application to start processing user queries.

Future Enhancements

Knowledge Base Expansion: Continuously update the vector store with new data. Retrieval Optimization: Explore more efficient retrieval techniques. Multi-lingual Support: Enable the application to handle multiple languages. Integration with Other Applications: Integrate with other tools or platforms for broader use cases.