Scoped retrieval
Web evidence is temporary per request, while file chunks stay tied to a selected conversation until cleanup.
CASE STUDY / AI RESEARCH
AI Search
Engine.
A citation-aware research workspace for searching the live web and selected documents, then inspecting the evidence behind each answer.
OVERVIEW
Research that keeps its sources close.
AI Search Engine brings live Tavily retrieval and selected PDF, TXT, Markdown, or DOCX documents into one conversation. Answers stream progressively and keep their own source list so citations remain inspectable over time.
ARCHITECTURE
One answer model, separate evidence lifecycles.
Frontend
React/Vite provides the workspace, local IndexedDB history, deep links, and streamed rendering.
Backend
FastAPI handles search, answer, streaming, upload, document deletion, and orchestration.
Web evidence
Tavily search → fetch/extract → chunk → request-scoped Qdrant retrieval → cleanup.
File evidence
Selected conversation documents are extracted, persistently indexed, and retrieved by scope.
Generation
Citation-aware context is passed to an Ollama provider; sources return with the answer.
Sources workspace
Per-message Web and File sources are available in a desktop panel or mobile drawer.
TECHNICAL DECISIONS
Citations as product UI
Inline citation markers connect to preserved source lists, making evidence easier to inspect.
Streaming by default
POST + fetch with ReadableStream SSE parsing renders progress and answer deltas as they arrive.
Honest evaluation
The deterministic offline quality gate reports 29/29 passes; citation metrics are structural regression checks, not factual correctness claims.
TRY IT
Explore the workspace.
Production runs on a Vercel frontend with a FastAPI backend on Render. The free-tier backend may cold-start after inactivity.