NexusIQ is a question-and-answer layer over your organization's own documents.

Find the answer without finding the file.

For the engineer chasing a decision made two years ago, the presales lead answering an RFP by Friday, and everyone who asks a colleague because search only hands back files.

EngineeringPartner Platform7 indexed · 1 pending

Which authentication flow did we settle on for the partner API, and why?

The team settled on OAuth 2.0 authorization code with PKCE, because partner mobile apps cannot hold a client secret safely. Access tokens carry a fifteen-minute TTL and refresh tokens rotate on every use, so a replayed token is revoked server-side.

Generated from 2 retrieved passages. Nothing outside them was used.

Cited passages

  • ADR-014-Partner-API-Auth.md§3 · Decision
    We adopt OAuth 2.0 authorization code with PKCE. Client credentials were rejected because partner mobile apps cannot hold a secret safely.
  • Security-Architecture-v4.pdfp. 12 · Session handling
    Access tokens are issued with a fifteen-minute TTL. Refresh tokens rotate on every use and are revoked server-side when a replay is detected.
Generated on infrastructure the operator controls, from the passages above and nothing else.
Runs on your servers
The whole system installs inside your own network. A hosted option exists, but your own infrastructure is the default.
AI on infrastructure you control
Generation, embedding and reranking run on a server the operator owns - self-hosted, or through your own gateway. It ships that way, and a deployment can be pinned so no administrator can point it at a third-party model.
Every answer shows its source
Each answer carries numbered references. Open one and the exact paragraph appears in the document it came from.
Publish it as a chatbot
A notebook becomes a chatbot on a link - public, passcode-protected, or organization members only - and embeds in any website as an iframe or a floating button.

The problem

The answer exists. Reaching it costs an afternoon.

Three moments out of one ordinary week, in three different jobs. None of them looks like a search problem, and all of them end the same way: somebody asks a colleague, and the colleague answers from memory.

  1. Engineering and PM

    A new engineer asks which auth flow the team settled on.

    The decision was made two years ago and written into an architecture record that sits in a repository nobody has opened since the handover. Three people answer from memory, one of them is wrong, and the flow gets re-argued in a meeting that did not need to happen.

    An office desk buried under loose paper, open notebooks and sticky notes.
  2. Sales, presales and BA

    An RFP lands on Tuesday and wants a comparable project by Friday.

    The case study exists. It is a deck in somebody's shared folder, saved under a file name that never mentions the client's industry, next to a capability statement that answers the compliance question further down. Finding both takes longer than writing the response.

    A person at a computer searching through a library catalogue.
  3. Everyone else

    Somebody needs the current rule, not the forty-page policy.

    The policy is on the intranet, it was updated in March, and the sentence they need is on page eight. They ask in the team channel instead and get an answer that was true last year. The same question comes back next month from somebody else.

    An office worker sitting behind a tall stack of paper documents, stamping them one sheet at a time.
This is what search gives them: the question typed once, six file names back, and an empty space where the answer was supposed to be. So it gets asked again next week, and the week after.

And the obvious fix uploads all of it.

The usual answer to this is an AI service that reads your document store in somebody else's data centre. For a bank, a hospital or a defence supplier, that is where the evaluation ends. NexusIQ runs the index and both models inside your network, so the documents and the question never leave it.

Documents lifting off a desk and floating up into a storage box that sits on a cloud.

Four jobs, four kinds of question

NexusIQ is used by whoever needs a fact out of the company's own material. Pick a role to see the question it gets asked and the answer it hands back.

They do not want the document. They want the one sentence inside it that settles the question, and proof that the sentence is the current one.

OperationsCompany policies7 indexed · 1 pending

Can I expense a client dinner in Tokyo, and what is the limit per head?

Client entertainment is reimbursable when a line manager approves it in advance, and the cap in Japan is 15,000 JPY per head. A claim above the cap needs a written justification attached, or finance sends it back.

Generated from 2 retrieved passages. Nothing outside them was used.

Ask anything about the sources in this notebook

Cited passages

  • Expense-Policy-2026-03.pdfp. 8 · Client entertainment
    Client entertainment is reimbursable when approved in advance by a line manager. The per-head cap is 15,000 JPY in Japan and 1,200,000 VND in Viet Nam.
  • Finance-FAQ.md§2 · Above-cap claims
    A claim above the per-head cap is accepted only with a written justification attached. Finance returns claims that arrive without one.
Generated on infrastructure the operator controls, from the passages above and nothing else.

The questions, answers and file names above are examples, rendered in the product's own interface. They are not customer material and not measured results.

Getting started

Three steps. No training day.

There is no query language to learn, no tagging exercise and no admin project to finish first. Somebody who can use a search box can use this on their first afternoon.

  1. Step 1

    Put the documents in

    Drag files into a notebook, paste a URL to crawl, or connect Google Drive. A source is pending while it is being indexed and ready afterwards - the interface says which, rather than leaving you guessing.

  2. Step 2

    Ask in your own words

    Type the question the way you would ask a colleague. Numbered citations come back with the answer and open the passage they came from. Ask for it concise, detailed or step by step if you want it shaped a certain way.

  3. Step 3

    Take the result with you

    Save the answer as a Studio output - a table, a mind map, a deck, a report - fix it where it is wrong, and put it back into the notebook as a source. Or publish the whole notebook as a chatbot.

All three happen inside the same notebook. Nothing has to be exported, reformatted or uploaded again to move between them.

A chat reply is the floor, not the deliverable

The Studio turns a notebook's sources into something you can hand on: twelve output types, generated on demand from the same cited corpus.

  • Summary

    A condensed read of the notebook's sources, for somebody who will not open them.

  • Mind map

    One topic decomposed into a hierarchy you can collapse branch by branch.

  • Knowledge graph

    The things named across the sources and how they connect, as a graph you can walk.

  • Data table

    Values pulled out of prose into rows and columns you can sort and export.

  • Slide deck

    A deck structure, section by section and point by point, ready to build on.

  • Report

    A longer structured document assembled from the notebook's sources.

  • Note

    A short written piece you keep working on inside the notebook.

  • Q&A

    The questions a reader of the sources would ask, answered and cited one by one.

  • FAQ

    Those same questions grouped by topic, collapsed until you open the one you need.

  • Flashcards

    A fact on the front, the answer on the back, ready to flip through and review.

  • Quiz

    Multiple-choice questions from the sources, with the correct answer marked and explained.

  • Statistics

    Key figures pulled into cards and charts, with anything the sources do not state marked as a gap.

  • The point is the artifact

    A chat reply is the floor. What a colleague receives is the table, the deck or the report - and it still points at the sources it was built from.

Editable in place
An output is a draft you correct, not a black box you regenerate and hope. Fix the wording where it is wrong and keep the rest.
Saved back as a source
A finished output can be written back into the notebook as a new source, so the next question can retrieve and cite it.
You are told when it is ready
A running output is tracked while you keep working. It survives a reload and a change of screen, and a notification says when it finished or failed.

Publishing

Turn a notebook into a chatbot the rest of the world can use

A chatbot is a thin publishing layer over exactly one notebook. It never owns its own sources, so a document added to the notebook is in the bot immediately, and access control has one implementation rather than two.

Who the link admits

  • Anyone with the link

    A public link, for a help page, a product site or a partner portal. Going public takes two deliberate acts, not one switch.

  • The link plus a passcode

    For a customer or a supplier. Wrong attempts are counted and the link locks itself - five tries and fifteen minutes by default, set by your organization.

  • Organization members only

    Signed-in members of your own organization. Never a candidate, never a customer, even if the link is forwarded.

The rules, which are boundaries rather than defaults

  • Two acts to go public. Publish the bot, then set its visibility to public. Unpublishing forces visibility back to the workspace.

  • Every link expires. Ninety days at the very most. There is no permanent link to lose track of.

  • Revocation is immediate. Access is re-checked on every request, so an open session is never trusted on its own.

  • A guest sees the passage, not the file. The answer, the citation marker and the cited excerpt - never a file name unless you enable it, never a URL, never a download.

  • A bot that cannot cite is not shippable. Publishing is refused when the set of sources it is allowed to read comes to nothing.

On someone else's website

www.example.co.jp/support
Support AssistantAnswers from your handbook

How long is the warranty on a replacement unit?

A replacement unit carries the remainder of the original warranty, or ninety days from the replacement date, whichever is longer.

“…where a unit is replaced under warranty, the replacement is covered for the remainder of the original term or ninety days, whichever is longer.”

The guest sees the passage, not the document it came from.

A drawing of the embedded bot, not a live one. What it shows - the numbered marker, the excerpt, the absence of a file name - is what a guest actually gets.
Two shapes, one builder
A plain iframe of the bot's page, or a floating launcher added with one script tag. Position, size, accent and label are set in the builder and copied out as the snippet.
A list of sites that may frame it
Each bot carries its own allow-list of origins, enforced by the browser as frame-ancestors. Leave it empty and any site may embed it; fill it in and only those may.

Point it at the systems your documents already live in

Sources arrive three ways: a file, a URL, or a system you already run. The catalogue lists 22 systems. Google Drive connects off the shelf today; every other row is built for an organization that asks. Whichever route a document takes, the chunking, the embedding and the index stay on your infrastructure.

  • File uploadDocuments and spreadsheets uploaded straight into a notebook.
  • URL crawlA web page or internal site fetched by the crawler and indexed as a source.

Cloud storage7

Business apps4

Project management3

Knowledge base2

Messaging4

Developer2

Vendor names stay in their own spelling. A connector reads the documents your account can already see - it grants nobody new access. No date is promised for an on-demand connector: ask, and we scope it with you.

Four stages, all of them on hardware you control

A question runs through your own index and a model on infrastructure the operator controls, and comes back pointing at the passages it used.

  1. Step 1

    Sources enter a notebook

    Upload files, point at a URL to crawl, or connect a system you already run. A workspace also holds a shared library every notebook in it can draw on.

    Office documents pass through a conversion service, web pages through a crawler.

  2. Step 2

    Chunking and embedding, in place

    Each source is split into chunks and embedded by an embedding model on infrastructure the operator controls. Ingestion is asynchronous: a source is pending first, then indexed.

    Vectors are written into pgvector at 1024 dimensions per chunk.

  3. Step 3

    Semantic search, then reranking

    A question runs semantic search across the notebook's private chunks and the workspace library, and the candidate passages are reranked before a word is generated.

    Retrieval follows permissions: a chunk you cannot open is a chunk you cannot retrieve.

  4. Step 4

    Generation grounded in the retrieved passages

    The generation model writes the answer from those passages and streams it token by token. Each numbered citation opens the passage it came from in the document viewer.

    Two paths ship, both on infrastructure the operator controls: a self-hosted gateway by default, and fully local models where there is no network path at all.

That is the whole mechanism. A cloud-only product can match the first and last stage; it cannot run the middle two on your side of the firewall.

Security

What is built, what your infrastructure provides, and what is not built yet

These three lists come from the engineering team's own security document, which sorts every control into one of them. A control that has not been built does not get to sit in the first column because it would read better there.

Built into the product

In the codebase, demonstrable against a running deployment.

  • Tamper-evident audit log

    Authentication and privileged events are hash-chained in the database. Updates and deletes are refused by the database itself, and the whole chain can be re-verified on demand.

  • An audit log the organization can read

    Owners and admins see their own organization’s timeline - actor, event, time, coarse location - and export it as CSV. The location is looked up from an offline database; no address is sent to another company for it.

  • Row-level security in PostgreSQL

    Organization-scoped tables carry database policies as well as application filters, so a query that forgets its organization returns nothing instead of another tenant’s rows.

  • Short-lived RS256 tokens

    An access token is signed RS256 and lives fifteen minutes. The session rides on an opaque refresh token that rotates on every use; replaying a spent one revokes the whole family.

  • TOTP multi-factor authentication

    Supported per user, with single-use recovery codes shown once, replay refused inside a code’s own window, and the second factor placed before a session is issued on every sign-in path. It is not a phishing-resistant factor, and there is no SMS and no push.

  • Stored credentials encrypted

    Third-party tokens and API keys are encrypted with AES-256-GCM. Passwords are hashed with bcrypt, and an account locks after repeated failures. Document text itself is not encrypted - retrieval has to read it.

  • No document text on the inference path

    The gateway writes one metadata line per request through a field allow-list - no prompt text and no document text - and no deployment ever learns the inference server’s address or key.

  • Dependency and static analysis on every change

    The pipeline runs a dependency audit and a static analyser on every merge request, and neither job is allowed to fail quietly. Known advisories in the locked dependency set are tracked openly rather than claimed away.

Your infrastructure provides it

Real, and the customer's to supply. The product requires and documents it rather than implementing it.

  • Encryption at rest

    The product writes to whatever storage you give it. Provision encrypted volumes, or turn on your storage layer’s own encryption.

  • TLS termination

    The API speaks plain HTTP behind your reverse proxy. You choose the TLS version and the cipher suite, which is why nobody can state them on your behalf.

  • Network-level IP restriction

    An application-level allow-list exists as an operator setting; a firewall in front of the deployment is the stronger and more usual control.

Planned, not present

Buildable, not built, and deliberately undated.

  • Single sign-on

    SAML, OIDC and LDAP are designed and not built. No date is attached to them.

  • Per-organization IP allow-list

    The enforcement point exists; promoting it into a tenant setting with an admin screen does not.

  • Wider audit coverage

    The chain holds whatever is written to it. Permission changes, membership changes and the share-link lifecycle are not written to it yet.

  • Personal data export

    Not available. The endpoint says so plainly rather than promising an archive that nothing would send.

There is no SOC 2, no ISO 27001, no certification badge and no third-party penetration test on this page, in any wording - not as a badge, not as “in progress”, not as “aligned with”. Each of those needs an auditor, a certification body or a testing firm, and none has been engaged. No patching window is published either, for the same reason.

Deployment, access and limits

Where it runs, who can see what, and what we will not claim

Enterprise buyers read landing pages defensively, so here is the part usually saved for a follow-up call.

Where it runs

Inside the lineServicesDatabaseVector indexGeneration modelEmbedding model

Your network

Services, database, index and models all run inside your network. Document content never crosses the boundary, because nothing sits on the other side of it.

Hover, tap or focus any layer to read what it holds.

Who can see what

Authorization is granted at four levels - platform, organization, workspace and notebook. These are the five roles the platform level defines. Point at one to see the layer it acts on.

Tamper-evident audit log

Authentication and privileged events are written to an append-only, hash-chained log that the database itself refuses to edit or delete. Owners and admins read their own organization’s timeline and export it as CSV.

Before you try it

Up front

There are no benchmark numbers, customer logos, case studies or certifications on this page yet, and four practical things shape what a first trial looks like.

Bounded by a local model
Generation quality is set by the model you host, not by a frontier API. That is the trade you make for sovereignty.
Cold starts are slow
A model that has not been loaded can take up to ninety seconds to answer the first question. The interface says so instead of pretending it failed.
Studio outputs take time
An output is generated inside the request, so tens of seconds is normal for a long one.
Indexing is not instant
A new source is pending before it is indexed, and it cannot be retrieved until it is.

Ask it a question from your own corpus

The honest evaluation is a notebook of your own documents, one real question, and a look at whether the cited passage actually says what the answer says.

Built by Tinhvan Japan Co., Ltd. On-premise deployments are scoped with you before anything is installed.

What happens after you send

  1. Your message reaches the NexusIQ sales team at Tinhvan Japan. No ticket queue, no auto-responder.
  2. A person who knows the product replies within one business day, in the language you wrote in.
  3. If it fits, we scope a pilot on your own documents before anything is installed.

Prefer your own mail client?

tvj.sale@tinhvan.com

Talk to us

Tell us what your team keeps looking for

A sales engineer reads every message. Describe what you are trying to find and where it lives today, and you get a specific answer back - not a brochure.

0 / 2000

We use your address to answer you, and for nothing else.