Meridian
Retrieval at fleet scale.
A retrieval pipeline that holds up against twelve million documents and a tribunal-grade audit.
✺ — The problem
Meridian's first-gen RAG was answering legal questions with the right vibe and the wrong citations. Their largest customer threatened to walk after a partner relied on a fabricated paragraph during a deposition. The team needed retrieval that was correct, attributable, and fast — not impressive.
Sector
Legal tech
Year
2025
Duration
9 weeks
Team
1 Principal · 2 Engineers
Stack
✺ — Approach
The same arc as every engagement — tuned to this problem.
Define · What 'correct' means
We co-wrote the evaluation rubric with two senior associates. Forty test queries with gold-standard passages. Anything not grounded in a real citation was scored zero, no partial credit.
Build · Hybrid retrieval, reranked
BM25 + dense embeddings + structured filters, fused with a cross-encoder rerank. Citations carried through every layer — the model could not write a sentence without a source it could point to.
Operate · Drift evals in CI
The evaluation suite runs on every PR. Silent regressions in retrieval recall now block merges instead of reaching customers two months later.
✺ — Outcome
Three numbers we’d defend in public.
94%
citation accuracy on the held-out audit set
320ms
median end-to-end retrieval latency
0
fabricated citations in 6 months of production
“They didn't try to sell us a vector database. They asked us what 'correct' meant in our world, wrote a test suite around it, and then built the smallest thing that passed.”
CTO, Meridian