02Case study

Enterprise RAG Platform — MEDZ

A web platform developed during a two-month internship: a client-facing interface, a PostgreSQL data layer, geographic visualization, and two AI assistants — technical and navigation — backed by a retrieval-augmented generation pipeline with bi-encoder retrieval, cross-encoder reranking and Gemma 2.

Context: Internship · 2 months
Role: Individual implementation under professional supervision
Timeline: July 2026 — September 2026
Stack: Next.js · React · PostgreSQL · pgvector · Prisma · LangChain · BGE-M3 · Cross-encoder reranking · Gemma 2 · Ollama

Screenshots GitHub

01 — Overview

Overview

During a two-month internship, I redesigned and developed MEDZ's web platform and integrated two AI assistants into it. The work was an individual implementation carried out under professional supervision.

Professional, client-facing web interface.
PostgreSQL database managed with Prisma ORM.
Mapping / geographic visualization.
A technical AI assistant and a navigation assistant built on retrieval-augmented generation.

MED-Z works in Design, development, commercialization, and management of business and industrial parks

02 — Problem

Problem

after long discussions with the company we identified some weaknesses in the existing platform and user needs for the AI assistants.

03 — Architecture

Architecture

User query
EmbeddingBGE-M3 (bi-encoder)
PostgreSQL + pgvectorSemantic retrieval
Top 25 candidates
Cross-encoder reranking
Top 5 contexts
Gemma 2Served with Ollama
Generated answer

Retrieval runs in two stages. A bi-encoder (BGE-M3) embeds the query and retrieves a broad candidate set from PostgreSQL with pgvector. A cross-encoder then rescores those candidates jointly with the query, and only the highest-ranked contexts are passed to Gemma 2 for generation.

04 — Data / Inputs

Data / Inputs

Document used in chunking was generated manually that explain the insights and processes involved and also document that explain the navigation and guidance of using the web-plateforme

05 — Methodology

Methodology

01EmbedQueries and documents are embedded with BGE-M3.
02RetrieveSemantic search over pgvector returns the top 25 candidate chunks.
03RerankA cross-encoder rescores the candidates against the query and keeps the top 5.
04GenerateGemma 2, served through Ollama, generates the answer from the selected contexts. LangChain wires the pipeline together.

The two-stage design trades a small amount of latency for precision: the bi-encoder is fast enough to search the whole index, while the cross-encoder is more accurate but only practical on a short list.

the bi-encoder used is BGE-M3 using cosine similarity and the cross-encoder was BERT.

06 — Engineering Implementation

Engineering Implementation

Next.js and React for the web platform and assistant interfaces.
PostgreSQL with pgvector as a single store for application data and embeddings.
Prisma ORM for schema and data access.
LangChain for the retrieval and generation pipeline.
Ollama for serving Gemma 2.

Mapping library type is Object-Relational Mapping (ORM) and the specific library used is Prisma.

07 — Evaluation / Results

Evaluation / Results

MEDZ web platform: Investir avec MEDZ page with navigation, hero section and the investment journey steps — Platform — client-facing interface (« Investir avec MEDZ » page)

pgAdmin query on the RagChunk table showing 123 rows with JSONB metadata: section order, estimated tokens, chunking strategy semantic_section — Knowledge base — RagChunk table in PostgreSQL (123 chunks, semantic-section chunking metadata)

pgAdmin view of the RagChunk table showing embeddedAt, createdAt, updatedAt and the pgvector embedding column — Knowledge base — stored embeddings (pgvector column) with timestamps

08 — Challenges & Trade-offs

Challenges & Trade-offs

the real challenges was to implement the solution based on both axes AI and development web .

09 — Resources

Resources

https://github.com/charfx/Medz_project.