Agents

Delegate the work, keep the judgement.

Agents that read your sources, use your tools, and show their work — so your team reviews decisions instead of assembling context.

Book a demo →
Ragenta — Agent dashboard

Agents

7 active

Today

Support triage — inbox #general

Retrieving·38 tickets

Vendor MSA review — Nordwind

Comparing clauses·64 clauses

RFP security questionnaire

Drafting gaps·112 / 118

Weekly engineering digest

Complete·5 themes

This week

Onboarding FAQ refresh

42 answers updated

Contract renewal watchlist

9 accounts flagged

Help-centre gap analysis

17 missing articles

⌘↵

Meet your team where they work

The same agents, the same knowledge, wherever the question is asked.

Web app

A workspace for chat, sources, agents and evaluation in one place.

answer.ts

import { Ragenta } from "ragenta"

const r = await agent.ask(

"refund window?"

)

API & SDK

Call an agent from your own product with a few lines of code.

Ragenta
Ask about your account…

Embedded widget

Drop a grounded assistant into your app or help centre.

#ask-internal

Ragenta

Retention is 90 days on Team plans — see Security policy §4.

Slack & Teams

Answers in-thread, with the same citations and permissions.

It understands your organisation, not just your documents

Context, conventions and permissions carry into every run.

run · parallel

Contract terms

done · opus-4.5

Account history

running · gpt-5

Support tickets

running · gpt-5

Billing records

queued · qwen3

Parallel subagents

A broad question fans out across subagents, each answering one part and streaming findings back as it finishes.

retrieval · trace

“What did we commit to Nordwind on pricing?”

MSA v4§7.2 — renewal pricing
Notesno increase before FY27
Q3 memograndfathering rule
Order formsigned 12 Mar 2026

Knowledge with provenance

Every passage carries its source, its permissions and its freshness — so an answer can be audited, not just read.

workspace · rules

Never quote an unapproved price

all agents

Cite the contract, not the summary

legal, sales

Escalate below 0.8 groundedness

support

Reply in the customer's language

support

Rules your team writes once

Tone, mandatory disclaimers, forbidden sources, escalation thresholds — written once and applied to every future run.

The whole loop, in one system

Ingest, retrieve, evaluate — and the evaluation feeds back into the first two.

ingest · pipeline
1

Parse — layout-aware, tables preserved

2

Chunk — on structure, not character count

3

Embed — multilingual, one vector space

4

Sync — incremental, permissions carried

ConfigureRun

Ingest

Connect a source and Ragenta parses layout, chunks on structure, embeds and keeps it in sync — with the source's permissions attached.

retrieve · hybrid
Vectortop 50 by similarity
50
KeywordBM25, exact identifiers
24
Mergededuplicated candidates
38
Rerankabove relevance threshold
6

Retrieve

Vector similarity and keyword matching run together, then a reranker orders what actually reaches the model's context.

evaluate · run

score 30 reference questions

groundedness 0.94 (+0.06)

check citation coverage

98% of claims linked

diff against last configuration

2 answers changed

Weakest area: retrieval on scanned PDFs

Evaluate

Groundedness, citation coverage and retrieval hit rate are scored on every change, so a regression surfaces before a user reports it.

What an agent can actually do

Read, act, and leave a trail you can follow.

ragenta · run · sandboxed

$ retrieving across 4 sources…

reranked 38 → 6 passages

$ sql: revenue by region, last 4 quarters

answer drafted · 9 claims · 9 citations

Run tools, not just prompts

Query a warehouse, call an internal API, run Python in a sandbox, search the web — with credentials held server-side and never shown to the model.

compose · context
Compare @vendor-msa.pdf against our @legal-playbook and apply @risk-rules.md
Use only sources newer than 2026-01-01

Point at exactly what matters

Mention a document, a folder, a connector or a saved filter inline, and the agent scopes its retrieval to it for that run.

history · versions

Knowledge base re-indexed

24 Aug · 09:12

Reranker enabled

24 Aug · 11:40

Model switched to opus-4.5

25 Aug · 08:05

Current configuration

26 Aug · 14:22

Every run is reproducible

Model, prompt, retrieved passages and tool outputs are versioned together, so any answer can be replayed exactly as it was produced.

Extend it without forking it

Connectors, skills and your own tools — all first-class.

connectors · workspace
Google Drive
synced
Notion
synced
Confluence
synced
Slack
syncing
Zendesk
synced

Connectors

Managed sync for the systems your knowledge already lives in, with each source's permissions carried through to retrieval.

skills · shared

/support-reply

Draft a support reply grounded in policy

/contract-diff

Compare a contract against the playbook

/rfp-answer

Answer a security questionnaire from prior answers

/weekly-digest

Summarise a week of engineering activity

/clause-search

Find every document that mentions a clause

Agent Skills

Turn a prompt that works into a named, versioned skill with its sources, tools and output format fixed — then let the whole workspace run it.

mcp · servers
internal-crm
connected
warehouse-sql
connected
ticketing
connected
billing-api
connected
custom-search
connected

Your own tools

Attach any Model Context Protocol server, or register a plain HTTP tool, and it appears to your agents alongside the built-in ones.

From the teams running it every day

We had three separate search tools and none of them could answer a question. Ragenta replaced all three in a month, and the citations are the reason people actually trust it.

HO

Head of Operations

Logistics, 900 staff

Support deflection went from a slide in a strategy deck to a number on a dashboard. The retrieval trace is what let us debug it ourselves instead of filing tickets.

CS

Customer Support Lead

B2B SaaS

Permission-aware retrieval was the whole deal for us. Legal would not sign off on anything that indexed HR files into one bucket.

ID

IT Director

Financial services

Our documentation is half Vietnamese, half English. One index handles both and an English question still finds the Vietnamese paragraph.

EM

Engineering Manager

Fintech, Ho Chi Minh City

The agent does not just retrieve — it decides what to look up next. Watching it re-query after a weak first result was the moment this clicked for the team.

HO

Head of Data

Healthcare group

We swapped the underlying model twice in a quarter without touching a single prompt. That flexibility is worth more than any single model benchmark.

PE

Principal Engineer

Enterprise software

Recently shipped

1.624 Aug 2026Agent Skills — reusable playbooks for your whole workspace
19 Aug 2026Cohere Rerank 3 available on every plan
11 Aug 2026Google Drive and Notion connectors now sync permissions
1.530 Jul 2026Retrieval trace on every answer

Put an agent on your knowledge today

Book a demo