🔍 Mistral-Powered Web QA App (LangChain + OpenRouter)

This project is a lightweight yet powerful question-answering system that allows users to:

🧠 Ask questions about the content of any public web page URL, powered by the mistralai/mistral-small-3.2-24b-instruct model.

Live Demo Link - Huggingface

---

Demo Video -

LLAMAINDEX.mp4

Without LangChain, this would require 200+ lines of manual data handling and orchestration logic. With it, you built a smart, extensible RAG system in ~30 lines of code.

---

🚀 What It Does

• ✅ Takes a URL as input
• ✅ Loads and splits the page into chunks using LangChain
• ✅ Converts chunks into vector embeddings (MiniLM)
• ✅ Performs semantic retrieval using FAISS
• ✅ Uses Mistral 3.2 (24B, free via OpenRouter) to synthesize answers

---

🧠 Why It's Not Just "API Wrapping"

This isn’t just plugging into an API. The system includes:

• 📄 Document Parsing: URL fetching, text extraction, chunking logic
• 📊 Semantic Search: Uses HuggingFace embeddings + FAISS vector search
• 🔄 Retrieval-Augmented Generation (RAG): Uses retrieved text as context for LLM to answer accurately
• 🧩 LangChain Chains: Modular chaining logic to connect retriever + LLM

This pipeline simulates the behavior of a fine-tuned ML model for Q&A — without requiring training from scratch.

---

🛠️ Tech Stack

Component	Tool
Text Retrieval	LangChain (WebBaseLoader, RecursiveCharacterTextSplitter)
Embeddings	sentence-transformers/all-MiniLM-L6-v2
Vector Store	FAISS
LLM (Free API)	mistralai/mistral-small-3.2-24b-instruct:free
Frontend	Gradio
Deployment	Hugging Face Spaces

---

🧑‍💻 Use Cases

• Ask questions about documentation or articles
• Educational summary generation
• Build RAG apps without training your own LLM
• Great base for interview-prep bots, study assistants, etc.

---

🔐 API Key Management

To use this on Hugging Face Spaces, we store the OpenRouter API key securely via Space Secrets (ArjunHF).

Absolutely! Here's an expanded and clear explanation of LangChain’s role in your project, written in clean, professional Markdown, ideal for README.md or project documentation:

---

🔗 What is LangChain Doing in This Project?

LangChain acts as the orchestrator that connects different components—document loaders, chunking logic, embeddings, retrievers, and LLMs—into a single, smart pipeline.

Here’s how LangChain powers the entire flow:

---

📥 1. Document Loading

WebBaseLoader

• LangChain uses WebBaseLoader to fetch and clean the raw content from a given web URL.
• It abstracts away boilerplate scraping code.
• Returns a list of Document objects for downstream processing.

---

✂️ 2. Text Splitting

RecursiveCharacterTextSplitter

• Large documents are split into manageable overlapping text chunks.
• This improves LLM comprehension and retrieval granularity.
• LangChain handles chunk boundaries intelligently using recursion on characters, newlines, sentences, etc.

---

🧠 3. Embedding & Vector Store

HuggingFaceEmbeddings + FAISS

• Each chunk is converted into a dense vector using a pretrained embedding model.
• These embeddings are stored in a FAISS index via LangChain’s VectorStore interface.
• LangChain lets you use this vector store as a retriever later on.

---

🔍 4. Context Retrieval

retriever.as_retriever()

• When a user asks a question, LangChain performs semantic search over the FAISS index to find the most relevant chunks.
• These are passed as context to the LLM for more grounded answers.

---

🧠 5. Answer Generation

RetrievalQA chain

• LangChain uses a Retrieval-Augmented Generation (RAG) setup via RetrievalQA.from_chain_type().
• It plugs in the retriever + the OpenRouter-backed Mistral LLM.
• Automatically forms prompts like: "Given the context: <retrieved_docs> — answer the question: <user_question>"

---