A next-gen AI-powered assistant that thinks before it speaks — designed to reason, retrieve, and respond with unmatched accuracy.
🔗 Live Demo: Click here to experience it!
⚠️ Note: The app is hosted on a serverless instance. It may take up to 50 seconds to load on a cold start.
This project is a fusion of cutting-edge technologies to build a logic-first, hallucination-free AI assistant capable of intelligent Q&A over SHL knowledge base articles. It scrapes, understands, and reflects — before responding.
-
✅ SHL Knowledge Base Integration
Scrapes and structures SHL documentation into markdown usingcrawl4aifor optimal LLM parsing. -
🔍 Semantic Search with FAISS + HuggingFace
Uses high-quality embeddings for lightning-fast and meaningful document retrieval. -
🧩 Tuned Gemini LLM
Finely-tuned with handcrafted reasoning samples and backed by a markdown-powered system prompt to ensure accuracy and reflection. -
🌐 Live Link Attestation
Query inputs are scanned for links, which are scraped and processed on-the-fly to support answers with real data. -
🔁 Chained Query Architecture
Input → Vector Retrieval → Prompt Logic → Answer Generation — all handled seamlessly using LangChain. -
💻 Interactive UI
Deployed with Streamlit for a sleek, no-friction user experience. Cold starts may take up to 50 seconds (free Render tier).
| Metric | Score |
|---|---|
| HumanEval | 0.60 |
| LLMEval | 0.73 |
⚙️ Optimized for: Clarity, reflection, and precision in response generation.
- 🧠 LLM: Gemini (tuned with handcrafted samples)
- 🔎 Retrieval: FAISS + Hugging Face Transformers
- 🕸️ Scraping:
crawl4ai(HTML → Markdown) - 🧠 Prompting: Markdown system prompt w/ reasoning logic
- 🧱 Framework: LangChain
- 🌐 Deployment: Streamlit + Render
Ideas? Suggestions? Feel free to open an issue or drop a star ⭐ if you find this useful or inspiring!
This project is open-sourced under the MIT License.
Created with 💡, logic, and a sprinkle of LLM magic.