In episode 75, at 35:22, WisdomAI CPO Kapil Chhabra describes identifying intent, splitting the question, querying suitable sources, checking the returned information, and assembling an answer. Query decomposition is the splitting step; verification is additional work.
Different sources are optional. LlamaIndex’s worked example asks three subquestions about different periods in Paul Graham’s life, all against one essay index.
His customer example asks how many website visitors were subscribers. The agent retrieves visitor data through Google Analytics and subscriber records from BigQuery, then joins them. WisdomAI sponsored the interview; the customer result is the guest’s account, not an independently measured benchmark.
For a comparable workflow, our recommendation is to check the time window, user identifier, and subscriber definition before combining results. Answering two smaller questions does not establish that their results describe the same population.
RAG concerns retrieving evidence for generation. Decomposition can help plan several retrievals, but it does not replace checking the evidence or the final answer.
Hear it from the guest
“And then it needs to recompose the response to present it back to the user who is asking the question.”
Quotes lightly edited to remove filler words.
Sources
- Episode 75: Kapil Chhabra on decomposition — A multi-source implementation described by WisdomAI's CPO in a sponsored interview.
- LlamaIndex: Sub Question Query Engine — The worked example asks several subquestions over one essay, showing that decomposition need not require different sources.
