Query Decomposition

Decomposes queries where the answer to one sub-question is a prerequisite for formulating the next. The system must execute steps in a strict topological order.

• Mechanism: The output of Sub-Question 1 becomes an input parameter for Sub-Question 2.
• Example: "What was the box office revenue of the film directed by the person who won Best Director in 2020?" decomposes to: (1) Who won Best Director in 2020? → (2) What film did [Answer 1] direct? → (3) What was the box office revenue of [Answer 2]?
• Contrast: Differs from parallel decomposition where sub-questions have no interdependencies.

Splits a complex query into multiple independent sub-questions that can be executed concurrently to reduce total retrieval latency.

• Mechanism: A comparator question like "Compare the GDP of France and Germany in 2023" fans out to: (1) What is the GDP of France in 2023? and (2) What is the GDP of Germany in 2023?.
• Optimization: Results are aggregated only after all parallel branches return, using a synthesis step to fuse the independent facts.
• Key Metric: Reduces end-to-end latency from O(n) to O(1) retrieval depth.

A two-stage decomposition paradigm where the model first identifies the full set of sub-problems and then solves them in a curated sequence.

• Stage 1 (Decomposition): Prompt the model with "To answer this question, what simpler sub-questions must we solve first?"
• Stage 2 (Execution): Solve sub-problems sequentially, feeding prior answers into the context window for subsequent steps.
• Advantage: Prevents the model from attempting to solve an unsolvable intermediate step by forcing explicit planning before execution.

A method that tightly couples chain-of-thought generation with retrieval, using each reasoning sentence as a dynamic query to a knowledge corpus.

• Process: The model generates a partial rationale, immediately queries a retriever with that rationale, and conditions the next reasoning step on the retrieved evidence.
• Benefit: Eliminates the need for a single, perfect initial query by allowing the system to course-correct its search based on intermediate logical deductions.
• Contrast: Unlike static decomposition, IRCoT adapts the sub-question plan in real-time based on what is actually retrieved.

A programming framework that treats query decomposition pipelines as declarative modules and automatically tunes prompts and fine-tunes weights to maximize a target metric.

• Signature: Define a MultiHopQA module with input fields and output fields. The compiler generates an optimized decomposition strategy.
• Teleprompter: Uses few-shot examples to bootstrap and optimize the decomposition instructions without manual prompt engineering.
• Outcome: Produces a compiled program that outperforms hand-crafted decomposition prompts on held-out evaluation sets.

Parses a complex factual assertion into atomic, independently verifiable sub-claims to enable granular evidence retrieval.

• Example: "Tesla, founded by Elon Musk, delivered 1.8 million vehicles in 2023" decomposes to: (1) Tesla was founded by Elon Musk. (2) Tesla delivered 1.8 million vehicles in 2023.
• Verification: Each atomic claim is checked against a trusted knowledge base. The final verdict is a logical AND of all sub-claim verifications.
• Granularity: Enables pinpoint attribution of which specific part of a statement is false, rather than a binary true/false label.