High RAG Question 159 of 223

How do you implement multi-hop retrieval reliably?

GenAI / LLM · Speak this in 60–90 seconds · Faridabad & Delhi NCR

Text inTokens

TransformerAttention

Text outNext token

Decompose the question, retrieve per sub-question, and condition later queries on intermediate facts.

LLMs use tokens (not full words) because it helps them:

This is a process

question about RAG.

Panels listen for order,

trade-offs, and what you would actually do on a GenAI / LLM project - not buzzwords.

Stable token IDs

Each piece maps to a number the network can learn.

Fits the model

Fixed pieces are what transformers expect as input.

Before the model can read a sentence, it goes through these steps:

1
Decompose the question, retrieve
per sub-question, and condition later queries on intermediate facts.
2
Cap hops and verify
each hop against sources.
3
Naive one-shot embedding search
fails when no single chunk contains the full answer.
4
Context mix
Attention looks at nearby tokens together.
5
Next token
The model scores what should come next.
6
Decode
IDs turn back into readable text.

Let's see how a real sentence is tokenized (tokens may vary by model):

Input text

“Cap hops and verify each hop against sources.”

Tokenized output

Caphopsandverifyeachhop

Token IDs (example)

2987408337471632900

Note: Actual tokens and IDs depend on the tokenizer (e.g., GPT, Llama, etc.).

Decompose the question, retrieve per sub-question, and condition later queries on intermediate facts. Cap hops and verify each hop against sources.