How It Works
Exa pre-computes embeddings for speed. Google likely uses HyDE-style query expansion. Our pipeline combines both approaches with adaptive classification — choosing the cheapest extraction strategy per query, then compressing the output into structured triples. Every step is shown below.
Adaptive Retrieval Classifier
Before running the full pipeline, a lightweight LLM reads the page’s heading structure and your query to pick the cheapest extraction strategy. On a listicle, it returns headings directly. On a FAQ, it pulls the relevant section. On scattered prose, it runs the full embedding pipeline. The classifier badge shows which strategy was chosen and why.
Split HyDE (Hypothetical Document Embeddings)
Instead of matching your question directly, the system generates what the answer should look like, splits it into sentences, and embeds each separately. Document sentences are scored by their maximum similarity to any HyDE vector — surfacing passages that match different facets of the expected answer.
Semantic + Keyword Hybrid
Pure embedding similarity misses keyword matches. Pure keyword search misses meaning. Combining both — cosine similarity with n-gram boost — surfaces passages that are both semantically relevant and contain the right terms.
Clause Extraction + Semantic Triples
Full sentences contain noise. The system extracts informative clauses, then converts them into structured (subject, predicate, fact) triples via a cheap LLM call. Compare all three granularities and their token counts — triples are typically 40–60% smaller while preserving all grounding information.
See what AI actually reads on your site
Site Intelligence crawls your entire site and builds a knowledge graph of what AI models extract. See which claims survive, which get lost, and where your content contradicts itself.
Learn more View plans