How we built the Arcana synthesis pipeline

The hardest part is not retrieval

Most people assume that the hard part of building an AI search engine is finding the right sources. It is not. Finding sources is a solved problem. The hard part is turning ten partially relevant documents into one coherent, accurate, cited answer in under two seconds.

That is what the synthesis pipeline does. This post explains how it works.

Stage 1: Query decomposition

Before any retrieval happens, Arcana analyzes the query to understand its structure. Is it a factual lookup, a comparison, a causal question, or a request for a summary? Each query type has a different optimal retrieval strategy. A comparison question needs sources that address both sides. A causal question needs sources that establish mechanism, not just correlation.

This decomposition step adds roughly 80 milliseconds to total latency. It is worth every one of them.

Stage 2: Parallel retrieval

Once the query is decomposed, retrieval runs in parallel across multiple source strategies simultaneously. We fetch from news, from academic repositories, from official documentation, and from high-trust general web sources at the same time. Parallel retrieval is what makes sub-two-second responses possible.

Stage 3: Relevance scoring and deduplication

Raw retrieval returns more content than any synthesis step can usefully process. Before synthesis, each retrieved chunk is scored for relevance to the specific query and deduplicated against other retrieved chunks. We discard low-relevance and near-duplicate content aggressively. The synthesis step only sees material that passed both filters.

Stage 4: Grounded synthesis

The synthesis step takes the filtered, scored chunks and generates a response. Critically, the synthesis model is constrained to ground every claim in the retrieved material. It cannot add information that was not present in the retrieval set. This constraint is what makes citation possible and what prevents hallucination.

Stage 5: Citation assignment

After synthesis, a separate pass assigns citations to each claim in the response. This is not a simple lookup. The citation assignment model identifies which specific source passage supports each specific claim, not just which source was generally relevant. This is what enables claim-level rather than document-level citations.