Back to portfolio
Selected Work

Featured Projects

Selected builds spanning agentic AI, retrieval systems, and business intelligence dashboards.

FILE/01

QA System Using LlamaIndex & Google Gemini (RAG)

Built and documented a production-grade RAG system with dense vector embeddings and semantic search, reducing hallucinations and delivering accurate, context-grounded answers across multi-format text corpora.

Python LlamaIndex Google Gemini Vector Embeddings
Architecture
RAG QA system architecture using LlamaIndex and Google Gemini Offline ingestion pipeline chunks documents, generates embeddings, and builds a vector index. Online query pipeline embeds the user question, retrieves matching chunks, assembles context, and passes it to Gemini for grounded answer generation. OFFLINE INGESTION Document corpus PDFs and multi-format text Chunking & parsing LlamaIndex node parsing Embedding generation Google Gemini embeddings Vector index LlamaIndex vector store ONLINE QUERY User query Semantic retrieval Top-k relevant chunks Context assembly Query + retrieved chunks Gemini LLM generation Grounded response synthesis Grounded answer

System Architecture Breakdown

Part 1: Offline Ingestion Pipeline (Preparation Phase)
  1. Document Corpus: The raw foundation of the knowledge base, comprising unstructured or semi-structured data such as PDFs and multi-format text files. Serves as the single source of truth that the QA system will reference to ground its answers.
  2. Chunking & Parsing: Powered by LlamaIndex node parsing. Raw documents are split into smaller, manageable text segments (nodes or chunks), cleanly preserving structural elements (like headings, paragraphs, and tables) so context isn't broken mid-sentence.
  3. Embedding Generation: Powered by Google Gemini embeddings. Each parsed text chunk is transformed into a high-dimensional dense vector, capturing semantic meaning to enable conceptual matching over exact keyword strings.
  4. Vector Index: Managed via LlamaIndex vector store. Generated vectors and their associated text chunks are organized and stored in an index optimized for high-speed similarity searches (such as cosine similarity or Euclidean distance).
Part 2: Online Query Pipeline (Execution Phase)
  1. Semantic Retrieval: Triggered when a user query enters the system. The query is converted into an embedding and matched against the vector index via similarity metrics to retrieve the top-$k$ most relevant context chunks.
  2. Context Assembly: Combines the raw user query and the retrieved top-$k$ chunks into a structured prompt package, bridging retrieval and generation for seamless consumption by the Large Language Model.
  3. Gemini LLM Generation: Powered by the Gemini LLM. The assembled context and query are evaluated to perform grounded response synthesis, relying strictly on retrieved documentation rather than parametric hallucination.
  4. Grounded Answer: The final system output delivered to the user, providing a precise, verifiable, and context-backed response.
Get in touch

Contact

Have a role, a project, or a question? Fill out the form or reach out directly.

Email akbarhusainc29@gmail.com LinkedIn akbarhusain429 GitHub Akbarhusain429 Twitter / X Akbarhusai429