Retrieval pipelines
Hybrid, graph, rerank.
For systems where being right matters more than being fast, retrieval is most of the engineering. We design hybrid retrieval — keyword, dense, graph — with reranking and citation tracking that make answers checkable.
Why retrieval beats bigger models for most enterprise work
A 70B model that hallucinates is worse than a 7B model that quotes its sources. In domains where customers will eventually fact-check the system — legal, medical, finance, ops — retrieval architecture is the lever that moves trust, and trust is the metric that moves the business.
What we build into retrieval pipelines
Hybrid index
Keyword (BM25), dense (embeddings), and metadata filters scored together.
Graph context
Knowledge-graph traversal for entity-heavy domains where structure matters.
Reranker
Cross-encoder or specialized reranker tuned on your relevance judgments.
Citation tracking
Every claim in the generated answer linked back to a passage operators can audit.
Cache layer
Query-level and embedding-level caching for the long tail of repeat questions.
Freshness
Incremental indexing, soft-delete, and source-of-truth contracts with upstream systems.