Distributed Systems & RAG

Cirix

Engineered an end-to-end Retrieval-Augmented Generation platform supporting parallel ingestion of multi-format documents (PDFs, web URLs, YouTube transcripts) with source-verifiable citations.

Cirix
WebsiteGitHub

Cirix

Cirix is a Retrieval-Augmented Generation (RAG) system built on LangChain's Expression Language (LCEL). It unifies unstructured content from PDFs, web pages, and YouTube video transcripts into a single ingestion pipeline, enabling natural language querying with source citations.

Overview

Extracting insights across fragmented media often requires switching between dedicated tools. Cirix unifies document parsing, web scraping, and video transcript extraction into a standardized ingestion flow. Content is chunked, vector-embedded, and stored in a vector index to provide context-grounded answers with exact source page, URL, or timestamp references.

Key Features

  • Unified Multi-Source Ingestion: Standardized data normalization for raw PDFs, website text, and pre-timestamped YouTube captions.
  • Sentence-Aware Chunking: Recursive text splitting designed to preserve semantic sentence boundaries during vectorization.
  • Context-Grounded RAG Pipeline: LangChain LCEL chain composing prompt templates, Gemini 2.5 Flash, and similarity retrievers.
  • Verifiable Citations: Every generated response explicitly cites original source document pages, URLs, or video timestamps.
  • Session Memory: In-memory conversation history management via RunnableWithMessageHistory scoped per session ID.

Tech Stack

  • Frontend: Next.js 15 (App Router), TypeScript, Tailwind CSS
  • Backend: FastAPI, Python, Uvicorn, Pydantic
  • AI / RAG Framework: LangChain (LCEL), Gemini 2.5 Flash, Gemini Embeddings (gemini-embedding-001)
  • Vector Database: ChromaDB (langchain-chroma)
  • Document Processing: PyMuPDF (fitz), Trafilatura, youtube-transcript-api
  • Hosting: Vercel (Frontend), Render (Backend API)