What is Query Transformation

Users often don’t know exactly what they want

To optimize Retrieval-Augmented Generation( RAG) “Query Transformation”

comes in play. We adjust the users query to an appropriate level of abstraction

User’s query maybe too abstract (“Kuch achha bana do.”) or too specific (“Kal subah 7:43 AM pe besan ka chilla banao, jisme exactly 1.5 green chilli ka paste ho, aur 2mm mota ho.”)
In Query Transformation we refine the query, making it more focused or broad enough to give better result

Example

Mom said: “Vha se Vo le aao”

If you try to directly execute the command you will end up with wrong thing, Instead

Think about the Query, analyse it

Know about the context

Where is mom?

If she’s in the kitchen preparing aamras, and there’s no sugar on the table, “Vha se vo le aao” probably means “Bring some sugar from the Top Shelf.”

Alternate Approach

You could ask “Mom, what exactly do you want”

but this is a little risky like mom may get angry

In technical way users get annoyed by follow-up questions, too many clarification may degrade user experience

Once you get the intended meaning you can execute the command and most probably get the best result

---

Now our goal is clear to give better anwer/output we have to make the query better

Which can be done by following techniques:

1. Parallel Query

2. Query Decomposition

3. Step Back

4. Hypothetical Document Embedding