Okay, a little experiment. Okay, I have been researching how LLMs work with vectors and how AI search works for a while now. But rather than me warble on for a while (and more than 5 mins!), I have tried an AI approach to explain the research ;)

In fairness, I did listen back to this, and personally I was quite impressed. The explanation about quantization methods was particularly useful.

If vectors, AI search and quantization mean nothing to you - take a listen :)

Hey, this is a free podcast. However, if you feel you want to support me then check out Patreon. I will have some more detailed deep dives for Patreon members as well as one-to-one sessions.

Or just buy a unicorn a coffee here!

Oh, and yes, I have ended up on YouTube (doesn’t everyone eventually?):

https://www.youtube.com/@justfifteenmins but don’t worry, no ugly face to worry about (yet!).

🧠 What RAG Actually Does

RAG enhances LLMs by retrieving relevant external information (e.g. from documents or databases) at query time, then feeding that into the prompt. This allows the LLM to answer with up-to-date or domain-specific knowledge without retraining.

💸 Is RAG Expensive?

Yes, it can be — especially if:

* You repeatedly reprocess large documents for every query.

* You use high token counts to include raw content in prompts.

* You rely on real-time parsing of files (e.g. PDFs or Excel) without preprocessing.

This is where vector storage and embedding optimization come in.

📦 Role of Vector Storage

Instead of reloading and reprocessing documents every time:

* Documents are chunked into smaller segments.

* Each chunk is converted into a vector embedding.

* These embeddings are stored in a vector database (e.g. FAISS, Pinecone, Weaviate).

* At query time, the user’s question is embedded and matched against stored vectors to retrieve relevant chunks.

This avoids reprocessing the original files and drastically reduces cost and latency

⚙️ Efficiency Strategies

Here’s how to make RAG more efficient:

Strategy

Description

Benefit

Vector Storage

Store precomputed embeddings

Avoids repeated parsing and embedding

ANN Indexing

Use Approximate Nearest Neighbor search

Fast retrieval from large datasets

Quantization

Compress embeddings (e.g. float8, int8)

Reduces memory footprint with minimal accuracy loss

Dimensionality Reduction

Use PCA or UMAP to reduce vector size

Speeds up search and lowers storage cost

Contextual Compression

Filter retrieved chunks before sending to LLM

Reduces token usage and cost

Design for UI/UX is obviously an art form, but can AI do as good a job or better? or as the case may well be, using AI to help with the tedious stuff?

UX Pilot

Figma

Hey, this is a free podcast. However, if you feel you want to support me then check out Patreon. I will have some more detailed deep dives for Patreon members as well as one-to-one sessions.

Or just buy a unicorn a coffee here!

Oh, and yes, I have ended up on YouTube (doesn’t everyone eventually?):

https://www.youtube.com/@justfifteenmins but don’t worry, no ugly face to worry about (yet!).