https://m101yosef.github.io/tags/rag--llms/RAG & LLMsHugo Blox Builder (https://hugoblox.com)en-usFri, 01 Nov 2024 00:00:00 +0000https://m101yosef.github.io/media/icon_hu_23bc55e615685232.pnghttps://m101yosef.github.io/tags/rag--llms/https://m101yosef.github.io/project/metrwnos-chatbot/Fri, 01 Nov 2024 00:00:00 +0000https://m101yosef.github.io/project/metrwnos-chatbot/<h2 id="the-problem">The problem</h2> <p>Most people who contact a law firm are not ready to hire a lawyer yet. They have a question: about a contract, a dispute, an inheritance, and they want a quick, understandable answer before committing to a paid consultation. For a small firm like Meter we Nos, answering those questions manually eats into billable time, and ignoring them loses potential clients.</p> <h2 id="why-a-chatbot-and-why-now-nov-2024">Why a chatbot, and why now (Nov 2024)</h2> <p>A year ago this would have been a hard sell. Generic chatbots gave generic answers, and no law firm can afford a bot that improvises legal advice. What changed in 2024 is that retrieval-augmented generation became practical and cheap: the model answers from the firm&rsquo;s own indexed legal documents instead of its imagination, and fast, low-cost models like Gemini Flash make per-query costs negligible.</p> <p>The business case is simple:</p> <ul> <li><strong>First contact runs 24/7.</strong> Prospective clients get an immediate answer in Egyptian Arabic instead of waiting for office hours.</li> <li><strong>Lawyers stop repeating themselves.</strong> The most common questions get consistent answers drawn from the firm&rsquo;s own material, freeing lawyer time for cases that actually need a lawyer.</li> <li><strong>Better-qualified consultations.</strong> By the time someone books a session, the basics are already covered, so the intake queue fills with real cases rather than FAQ traffic.</li> </ul> <h2 id="system--architecture">System &amp; Architecture</h2> <p>Built as a contract AI engineering project around a Retrieval-Augmented Generation (RAG) pipeline:</p> <ul> <li><strong>LLM &amp; Embeddings</strong>: Google Gemini (<code>gemini-1.5-flash</code>) through LangChain&rsquo;s <code>RetrievalQA</code>.</li> <li><strong>Vector Retrieval</strong>: FAISS index over a corpus of Egyptian legal documents.</li> </ul> <h2 id="what-i-did">What I did</h2> <ul> <li>Designed and implemented the RAG pipeline end to end: document indexing, vector retrieval, and answer generation.</li> <li>Wrote the system prompt so responses come back as clear Egyptian Arabic legal explanations, suitable for talking to clients rather than lawyers.</li> <li>Built the Streamlit prototype and the client intake flow around it.</li> </ul> <h2 id="demo-walkthrough">Demo Walkthrough</h2> <p>The demo below (<code>v0.3</code>) shows the chatbot answering Egyptian legal questions and walking through the intake flow:</p> <video controls width="100%" style="border-radius: 8px; margin-top: 1rem; box-shadow: 0 4px 12px rgba(0,0,0,0.15);"> <source src="demo.mp4" type="video/mp4"> Your browser does not support the video tag. </video> <h2 id="outcome">Outcome</h2> <p>Delivered a working prototype through three demo iterations (<code>v0.1</code> to <code>v0.3</code>), enough to show the firm that document-grounded LLMs can realistically handle legal intake for a boutique practice.</p>