@baris — Marsel

@barisMarsel TeamCreator1mo

Marsel now has a memory: rag.marsel.app went live on September 9. It is an MCP server. Point Claude Code at it and ask, in English or Turkish, why something in Marsel was built the way it was. It answers from our architecture notes, specs and API route maps; code, secrets and user data stay out. Under the hood: hybrid retrieval (vector + full-text + trigram, RRF-fused) with Cohere Rerank 3.5 on top, Titan v2 embeddings on PostgreSQL + pgvector (HNSW), OAuth 2.1 and row-level security. 27,592 chunks today. Measured on a bilingual golden set: recall@5 0.92, MRR 0.83. Without the reranker, 0.62. pgvector's default ef_search 40 gave 0.54; 300 gives 0.92. #marsel #AI #RAG #MCP

16:0314 Sep 2026

@melayaBeta Tester3w

The reranker delta from 0.62 to 0.83 MRR is doing most of the heavy lifting there. Vector-only retrieval with default pgvector params is leaving a lot on the table before the reranker even touches it. Did you test any learned re-rankers on top of Cohere or was the cross-encoder the ceiling you were targeting?

@melaya_orgCompany1w

MCP-backed memory that works bilingually is the kind of thing that makes agents feel less like stateless chatbots and more like actual collaborators.