Project 02 · Enterprise AI

QueryDoc

Production-grade AI document intelligence platform for highly accurate, hallucination-free Q&A using Retrieval-Augmented Generation.

View on GitHub ↗

QueryDoc Interface — Developer-Focused Workspace

Overview

QueryDoc is a production-grade, AI-powered document intelligence platform designed to ingest local files (PDFs, DOCX, TXT) and provide highly accurate, hallucination-free answers using Retrieval-Augmented Generation (RAG).

Unlike standard prototype chatbots, QueryDoc was architected to solve the "hidden 70%" of production AI problems — focusing heavily on retrieval precision, strict citations, and system observability rather than just LLM prompting.

System Demonstration

Full demonstration of the QueryDoc platform — showcasing document ingestion, hybrid vector/sparse retrieval, cross-encoder reranking, and grounded Q&A with strict citations.

Architecture

User Query Input ↓ Parallel Hybrid Search (FAISS Dense + BM25 Sparse) ↓ Reciprocal Rank Fusion (RRF) ↓ Cross-Encoder Semantic Reranking (ms-marco) ↓ LLM Prompting with Strict Citations ↓ Grounded Response Generation

Key Features

Hybrid Search Retrieval: Fuses FAISS vector similarity with BM25 sparse search using Reciprocal Rank Fusion to capture both semantic meaning and exact keywords.
Cross-Encoder Reranking: Re-orders the top hybrid results using an ms-marco model to surface the most contextually relevant data.
Strict Citation Enforcement: LLM is aggressively prompted to ground answers exclusively in retrieved context, virtually eliminating hallucinations.
Telemetry & Observability: Custom TelemetryTracer tracking latency, chunk count, token usage, and cost per request.
Automated Regression Evaluation: Gated by CI/CD GitHub Actions testing context retrieval accuracy and answer correctness on every push.
Developer-Focused Frontend: Dense React/Vite SPA with Recharts analytics and a 70/30 split view layout.

Platform Interface

Developer-Focused Workspace

70/30 split view with document preview and retrieval metadata panel.

Telemetry & Analytics

Recharts-powered dashboard visualizing query volume, users, and latency.

Chat Interface

Grounded conversational Q&A with inline source citations.

Document Management

Upload, process, and manage document collections.

Tech Stack

Frontend

React 18 Vite Framer Motion Recharts

Backend

Python FastAPI LangChain GitHub Actions

Database & Vector Store

FAISS Vector DB SQLite

AI & ML Engine

Llama 3.1 8B (Groq) PyTorch Cross-Encoders