What's new in Ragenta.

Agent Skills — reusable playbooks for your whole workspace

Turn a prompt that works into a named skill anyone on the team can invoke, with its tools, sources and output format fixed.

Cohere Rerank 3 available on every plan

Reranking moved from a Team-only feature to the default retrieval pipeline, with a measurable jump in citation accuracy on long documents.

Google Drive and Notion connectors now sync permissions

Retrieval filters by the source system's own access control, so a member only ever sees passages from files they can already open.

Retrieval trace on every answer

Open any answer to see every candidate passage, its score, and whether it made the final context — the fastest way to tell a retrieval failure from a generation failure.

• ▸Per-claim citation highlighting
• ▸Score breakdown for vector and keyword matching
• ▸One-click re-run with a different model

Hybrid search is now the default

Vector similarity combined with BM25 keyword matching, so exact part numbers, error codes and ticket IDs are found again.

Vietnamese retrieval quality

Unicode normalisation at ingestion plus a bigram keyword index lift recall noticeably on Vietnamese corpora and on mixed-language internal docs.

Parallel subagents

A research question now fans out across independent subagents, each working one sub-question and streaming its findings back as it finishes.

Model Context Protocol servers

Attach any MCP server and its tools appear to your agents alongside the built-in ones, with per-workspace credentials held server-side.

Evaluation runs on every configuration change

Point an evaluation set at a knowledge base and Ragenta re-scores groundedness, citation coverage and retrieval hit rate whenever the configuration changes.

Large PDF ingestion no longer stalls at 90%

Documents above roughly 400 pages could stall during the embedding stage when a single page produced an unusually long table. Fixed, with a regression test on the ingestion queue.