Gen-AI Developer Classroom notes 08/Mar/2026

Enterprise Vector Store

• Possible options being
  • pgvector
  • cloud vector storages
  • pinecone

PgVector

• Ensure docker desktop is running

• This is postgres vector extension.

• setting up pgvector database

Give me steps to setup pgvector on a linux instance

• Create a folder called as initdb and create a file called as 01-enable-vector.sql with following content

CREATE EXTENSION IF NOT EXISTS vector;

• Now cd into this folder and execute the following commands

docker pull pgvector/pgvector:pg16
docker run -d \
--name pgvector-db \
-e POSTGRES_USER=admin \
-e POSTGRES_PASSWORD=admin123 \
-e POSTGRES_DB=vectordb \
-v $(pwd)/initdb:/docker-entrypoint-initdb.d \
-p 5432:5432 \
pgvector/pgvector:pg16

• Auto-Enable pgvector When Container Starts (Best Practice)

Enterprise Version Recommended for retrieval

• We will define the retrieval pipeline in 4 layers

  • Layer 1: retrieval
    • pgvector
    • embedding
    • metaddata filters
    • MMR
  • Layer 2: ranking
    • top 15 initial
    • rerank to top 5
  • Layer 3: grounding
    • strict answer from context prompt
    • citations mandatory
  • Layer 4: governance
    • log retrieved chunk IDS
    • log policy titles/sections used

Search types

• Similarity
• MMR (Maximal Marginal Relevance). It tries to balance relevance with diversity

• MQR (MultiQueryRetriever)

• Refer Here for implementation based on mmr

We need to give demonstration of this RAG

• We need a ui. A quick way to bring up a simple ui for demos is
• Streamlit helloworld
• Refer Here for sample streamlit which we have built.