Ask a large language model for a market size and it will give you one — confidently, precisely, and quite possibly invented. Hallucination is the single biggest obstacle to using AI for research that real decisions depend on. It is also the problem we have spent the most engineering effort solving. Here is how our pipeline works, and why the architecture matters more than any individual model.
Why Models Hallucinate Numbers
Language models are trained to produce plausible text, and a specific number is more plausible-sounding than "I don't know." Left unconstrained, a model asked about a niche market will blend half-remembered training data into a figure that looks authoritative and traces to nothing. For casual use that is an annoyance; for market intelligence it is disqualifying.
Our Answer: Separate Retrieval From Writing
RabbitReport's core design decision is that the model that writes the report is never the source of the facts. The pipeline has two distinct stages:
Stage 1 — Live Research
A dedicated research stage performs real web searches at generation time and returns a structured data dump in which every fact carries a citation to a live source retrieved during that session. This stage operates under strict integrity rules: if no source is found for a data point, it must say "no publicly available data found" rather than estimate. Training-data memory is explicitly off-limits as a source of figures.
Stage 2 — Constrained Writing
A second stage transforms the verified research into the finished report — under hard constraints. It may only use figures that appear in the research data. If the research gives a range, the report keeps the range instead of inventing false precision. If the research hedges ("estimates suggest," "approximately"), the report preserves the hedge. Claims may only be attributed to a source if the research explicitly names that source for that claim.
Honest Uncertainty Is a Feature
A subtle but important principle: a report that says "approximately $400B" or "data not publicly available" is more trustworthy than one with a precise figure for everything. We instruct the writing stage to preserve uncertainty language deliberately, and every report carries a sources page so claims can be verified by following the citations.
What This Means in Practice
No system eliminates error entirely — sources themselves can be wrong, and judgment is still the reader's job. What the dual-stage architecture eliminates is the specific failure mode that makes raw LLM output unusable for research: confident numbers with no origin. Every figure in a RabbitReport traces back to a real document retrieved at generation time. That is the standard we believe all AI research should be held to.