In the world of Retrieval-Augmented Generation (RAG), most systems still treat document retrieval like a popularity contest — fetch the most relevant-looking text and hope the generator can stitch the answer together. But as any manager who has tried to merge three half-baked reports knows, relevance without completeness is a recipe for failure.
A new framework, Compositional Answer Retrieval (CAR), aims to fix that. Instead of asking a retrieval model to find a single “best” set of documents, CAR teaches it to think like a strategist: break the question into its components, retrieve for each, and then assemble the pieces into a coherent whole.
Why Compositionality Matters
Imagine you’re researching a market trend that requires combining regulatory updates from two countries, recent commodity price data, and a competitor’s product release timeline. A conventional RAG system might bring you 10 documents all about one country’s regulations, missing the other crucial pieces. CAR, by contrast, treats the task like a supply chain — decompose, source each part, integrate.
This is especially critical in multi-hop question answering, where the answer depends on facts scattered across unrelated sources. Benchmarks like MuSiQue, HotpotQA, and 2WikiMultihopQA have long exposed the weakness of “relevance-only” retrieval: the best single document is rarely the whole answer.
How CAR Works
Step 1 — Question Decomposition
- CAR uses a decomposition model to split a question into targeted sub-questions.
- Example: “Which authors won both the Hugo and Nebula awards for the same work?” → two sub-questions, one for Hugo winners, one for Nebula winners.
Step 2 — Targeted Retrieval
- Each sub-question drives an independent retrieval pass.
- Encourages diversity so no component of the answer is missed.
Step 3 — Evidence Assembly
- The collected evidence is merged and passed to a generator (e.g., an LLM) to produce the final answer.
Under the hood, CAR employs a coverage-oriented loss function and reinforcement learning signals based on answer completeness, not just retrieval accuracy.
Component | Traditional RAG | CAR Framework |
---|---|---|
Retrieval Target | Single best-ranked docs | Docs covering all answer parts |
Training Signal | Relevance labels | Answer coverage reward |
Weakness | May miss key evidence | Prioritizes completeness |
Business Relevance
For enterprise applications, CAR’s philosophy maps directly onto real-world information workflows:
- Compliance: Combining regulations from multiple jurisdictions.
- Business Intelligence: Merging financial, operational, and market data.
- Research & Development: Integrating findings from separate studies.
By forcing retrieval to plan for coverage, CAR turns RAG from a hopeful guesser into a methodical investigator — a change that could significantly improve AI’s reliability in high-stakes decisions.
Looking Ahead
The CAR approach invites a rethink of retrieval objectives: what if we trained every enterprise search and analytics tool to value complete answer construction over local relevance? In regulated industries, in strategy consulting, in scientific synthesis — this shift could mean the difference between half-answers and actionable intelligence.
Cognaptus: Automate the Present, Incubate the Future