Darrell Mortalla
← Back to featured systems

Llm Engineering

AI RAG Knowledge Assistant

Grounded answers from retrieved knowledge, with source attribution.

A retrieval-augmented generation system that loads and chunks source documents, embeds them with OpenAI, stores vectors in FAISS, retrieves relevant context, and generates grounded answers with source metadata through FastAPI and Streamlit.

View repository

Problem

Context

Large language models can generate fluent answers without being grounded in a trusted knowledge source, creating a need for retrieval, attribution, and explicit context boundaries.

Challenge

Build a modular RAG workflow that retrieves relevant document context before generation and exposes the result through usable application interfaces.

Objective

Create an end-to-end knowledge assistant that performs document ingestion, chunking, embedding, vector retrieval, grounded generation, source attribution, and application delivery through API and UI layers.

Architecture

Documents are loaded and chunked, embedded with OpenAI text-embedding-3-small, stored in a FAISS vector index, retrieved by similarity, injected into a grounded prompt, and answered with gpt-4o-mini through API and UI layers.

  1. 01

    Source documents

  2. 02

    Document chunks

  3. 03

    OpenAI embeddings

  4. 04

    FAISS vector index

  5. 05

    Top-k retrieved context

  6. 06

    Grounded prompt

  7. 07

    LLM-generated answer with source metadata

Document Ingestion

Loads source documents and prepares text for downstream chunking and retrieval.

  • Python
  • LangChain

Chunking Pipeline

Splits source text into retrievable units while preserving source metadata.

  • Python
  • LangChain

Embedding Layer

Transforms document chunks into vector representations for semantic retrieval.

  • OpenAI
  • text-embedding-3-small

Vector Store

Stores and retrieves document embeddings through a vector-store abstraction backed by FAISS.

  • FAISS

Retrieval Layer

Selects top-k relevant chunks to provide grounded context for generation.

  • FAISS
  • LangChain

Generation Layer

Constructs a grounded prompt and generates answers constrained to retrieved context.

  • OpenAI
  • gpt-4o-mini

API Layer

Exposes retrieval-augmented question answering through an HTTP interface.

  • FastAPI

User Interface

Provides an interactive front end for asking questions and reviewing grounded responses.

  • Streamlit

Technology

Language

  • Python 3.11

Ml Ai

  • OpenAI text-embedding-3-small
  • OpenAI gpt-4o-mini
  • LangChain
  • FAISS

Framework

  • FastAPI
  • Streamlit

Engineering Contribution

Built the complete RAG retrieval pipeline

Implemented document loading, chunking, embeddings, vector storage, top-k retrieval, and grounded answer generation as a modular workflow.

Added explicit grounding constraints

Constructed prompts that instruct the model to answer only from retrieved context rather than relying on unsupported external knowledge.

Preserved source attribution

Returned source and chunk metadata alongside generated answers so retrieved evidence remains inspectable.

Separated vector-store implementation behind an abstraction

Implemented FAISS through a vector-store abstraction so retrieval infrastructure is not tightly coupled to application logic.

Delivered API and interactive UI access

Exposed the RAG workflow through FastAPI and Streamlit for programmatic and interactive use.

Documented current system boundaries

Kept limitations explicit, including the local corpus, FAISS-only backend, heuristic grounding signal, and absence of authentication, streaming, formal evaluation, and production deployment.

Engineering Evidence

Architecture

End-to-End RAG Architecture

The system implements document loading, text chunking, embeddings, FAISS storage, semantic retrieval, grounded prompting, and answer generation.

Code

Semantic Vector Retrieval

Document chunks are embedded with text-embedding-3-small, stored in FAISS, and retrieved as top-k context for question answering.

Code

Grounded Answer Generation

The generation prompt instructs gpt-4o-mini to answer from retrieved context rather than inventing unsupported information.

Documentation

Source and Chunk Attribution

RAG responses include source metadata and chunk-level attribution for the retrieved material used during generation.

Architecture

Vector Store Abstraction

The current FAISS implementation is accessed through a vector-store abstraction rather than being embedded directly into higher-level application logic.

Code

FastAPI and Streamlit Interfaces

The system exposes retrieval-augmented question answering through both a FastAPI service and a Streamlit interface.

Metric

Heuristic Grounding Signal

The system includes a heuristic grounding signal to provide an additional indication of answer support, but it is not a calibrated evaluation metric.

Documentation

Documented Current Limitations

The current implementation uses a small local text corpus and FAISS backend, includes no authentication or streaming, does not yet include a formal evaluation framework, and is not presented as a production deployment.

What It Proves

Retrieval-augmented generation

Demonstrates the full RAG path from source documents through retrieval to grounded LLM generation.

LLM application engineering

Demonstrates integration of embeddings, vector search, generation, source attribution, API delivery, and an interactive UI.

Evidence-aware AI design

Demonstrates explicit grounding constraints, inspectable source metadata, and documented limits on what the current system can claim.