Build a RAG Chatbot for Your Documents

▎ What you'll build

You leave with the real thing - not notes.

✓A document ingestion script that loads PDFs and Markdown, chunks them, and embeds each chunk

✓A pgvector table in Postgres with an HNSW index for fast similarity search

✓A retrieval function that embeds a question and pulls the top-k relevant chunks

✓An answer pipeline that feeds retrieved chunks to an LLM and returns answers with inline source citations

✓A 'don't know' guardrail that refuses to answer when retrieval finds nothing relevant

▎ The 3 hours, block by block

Hands-on the whole way.

Block 1

Ingestion and embeddings

Load documents, chunk with recursive 512-token splitting - and why chunk size and overlap matter
Generate embeddings via an API
Store chunks, vectors, and metadata in Postgres

Block 2

pgvector and retrieval

Enable the vector extension and build an HNSW index
Run cosine-similarity queries and tune top-k and ef_search
Add metadata filtering for scoped retrieval

Block 3

Answer pipeline and grounding

Assemble the retrieval-augmented prompt and call the LLM from Python
Add inline citations back to source chunks
Add a relevance threshold so the bot says 'I don't know' instead of hallucinating

Who it's for

Software engineers who want to build RAG from the primitives, not just call a black-box framework
Founders validating an AI-on-your-docs feature before committing to a vendor
Backend developers who already run Postgres and want vector search without a separate database

What to bring

A laptop with Python 3.10+ and a code editor
Comfort reading and editing Python - functions, virtualenv, pip
Local Postgres 14+ (we install pgvector together) or a free Neon/Supabase instance
An LLM / embeddings API key (OpenAI or equivalent) - we send setup steps before the session

▎ By the end

What's true when you walk out.

✓A running chatbot that answers from your own documents, end to end, on your machine

✓You understand each stage - chunking, embedding, indexing, retrieval, generation - well enough to debug and tune it

✓You can decide when pgvector-in-Postgres is enough versus when a dedicated vector database is worth it

▎ Tools you'll touch

PythonPostgrespgvectorHNSW indexOpenAI embeddings + chat APIpsycopgrecursive chunking

▎ Who teaches

Your instructor

Workshop instructor

I've spent 22+ years building software for enterprises - full-stack apps, backend systems, and lately RAG pipelines and agentic AI solutions. I've shipped the hard stuff for big companies. These workshops are that experience, distilled into one hands-on room so you can ship your own.

▎ Questions

Before you sign up.

Do I need a separate vector database like Pinecone?

No. We use pgvector inside the Postgres you already run, which comfortably handles millions of vectors with fast queries - plenty for most products. We also cover the signals that tell you when to graduate to a dedicated store.

Will the chatbot make things up?

We build specifically against that - the bot answers only from retrieved chunks, cites its sources inline, and is set to say 'I don't know' when nothing relevant is found. Grounding reduces hallucination but does not eliminate it, so we cover the limits honestly.

How much Python do I need to know?

Enough to read functions, run a script, and install packages. You do not need prior ML or embeddings experience - we build the pipeline from scratch and explain each piece as we go.

Sat 19 Sep 2026 · 3 hours · 2 PM - 5 PM · Chennai · 10 seats. Drop your email and we'll tell you the moment booking opens.