Zouhour Bellamine

Hello

English 🇬🇧

✿
Loading
← All work

01 / AI research / Full-stack system

AI Research Copilot

Research. Retrieve. Reason. Remember.

A full-stack AI research workspace designed to help users explore the web, analyze documents, retrieve relevant knowledge, and build contextual answers with citations.

AI ResearchRAGLLMsWeb ResearchFull-Stack
research.workspaceLIVE
How is RAG changing research workflows?
searching the webretrieving documentsbuilding context
010203sources connected

Contextual answer with citations and retrieved evidence.

The problem

Research is fragmented.

Answering a meaningful research question often means moving between web sources, documents, notes, previous conversations, and reports. This project brings those workflows together in one AI-powered workspace.

Web+Documents+Memory+AI↓One Research Workspace

Capabilities

One workspace, several research paths.

01

Research mode

Question→Exa→Relevant sources→Cited answer

Routes current research questions through Exa and brings returned sources into the response.

02

Document mode

Document→Extraction→Chunks + embeddings→Semantic retrieval

Retrieves relevant document chunks instead of sending an uploaded file in full to the model.

03

Memory

Conversation→Useful context→Semantic memory→Future conversation

Useful context can be embedded, stored, and retrieved across sessions when it is relevant.

04

Report generation

Research→Structured report→PDF / DOCX→Export

Turns research and conversation outputs into structured PDF or DOCX exports.

Core architecture

Separate paths, shared intelligence.

Next.js Frontend↓FastAPI Backend↓AI Orchestration

Research mode

Exa Web Search↓Cited Sources

Document mode

Document Processing↓Chunks + Embeddings
pgvector Retrieval↓Gemini / LLM↓Contextual Response

Supabase PostgreSQL + pgvector · Supabase Storage · Persistent Sessions / Memory

Document retrieval

RAG makes documents queryable.

The system retrieves relevant document chunks for a question instead of sending the entire uploaded document to the model.

01

Upload document

02

Extract content

03

Split into chunks

04

Generate embeddings

05

Store vectors

06

Semantic search

07

Retrieve relevant chunks

08

Contextual answer

Grounded research

From a question to evidence-backed answers.

When Research Mode is enabled, the application routes the request to Exa and returns available sources to the interface.

Research QuestionExa Web SearchRelevant SourcesSource ContentLLM ReasoningAnswer + Citations

Conversational memory

Context that can carry forward.

Useful personal facts and conversation context can be embedded and stored for later semantic retrieval. It is a retrieval feature, not human-like memory.

Session 01

User shares useful context

↓
Semantic Memory

Embedded and stored

↓
Session 02

Relevant context can be retrieved

The user experience

A research conversation with a clear path.

01

Start a conversation

02

Choose research or documents

03

Ask a question

04

Retrieve evidence

05

Generate a contextual answer

06

Continue the conversation

07

Export the result

Technical contribution

Building the system behind the interface.

01

Frontend

Next.js · React · TypeScript

Responsive workspace, conversation history, controls, uploads, sources, and streaming interaction.

02

Backend

Python · FastAPI

API routes, AI orchestration, research routing, document processing, memory, and reports.

03

AI layer

Gemini · Exa

LLM responses combined with web research and source retrieval.

04

Retrieval

Supabase PostgreSQL · pgvector

Persistent data and semantic retrieval for documents, memory, and research context.

05

Storage

Supabase Storage

Uploaded documents and generated reports.

Engineering decisions

Each tool supports a deliberate boundary.

Why RAG?To ground document-based answers in retrieved content.

Why Exa?A dedicated web-research path with source retrieval and citations.

Why FastAPI?To separate AI orchestration and backend services from the frontend.

Why PostgreSQL + pgvector?To combine persistent application data with semantic vector retrieval.

Why Supabase?Database, vector storage, and file storage infrastructure in one system.

Deployment

A live, connected production architecture.

VercelNext.js frontend→RenderFastAPI backend→SupabasePostgreSQL · pgvector · Storage

See it in action

See the Copilot in action.

From research questions to document analysis, retrieval, sources, and contextual answers.

Technical ecosystem

AIGemini · RAG · Exa

BackendPython · FastAPI

FrontendNext.js · React · TypeScript

DataPostgreSQL · pgvector · Supabase

InfrastructureDocker · Vercel · Render

Project scope

Research should not mean switching between ten different tools.

One workspace. Multiple sources. Contextual answers.

Next project: Text-to-SQL Data Assistant