Corporate Knowledge RAG — your own documents, answerable in seconds.

Your specs, contracts, drawings and manuals already contain the answer. The problem is finding it. We build permissioned retrieval systems that answer in natural language and cite the source file — so nobody has to trust the model, they can check it.

~150×
Faster document search
~500
Internal users served
TB
Indexed corpus scale
1
Permissioned source of truth

Measured on a live deployment — see the UK construction group case study.

The knowledge exists. Nobody can retrieve it.

In a document-heavy business, most of the working day is spent locating information rather than using it. On the UK construction deployment, estimators were spending roughly 15 minutes per lookup hunting through a terabyte-scale archive of specifications, drawings and contracts.

Full-text search doesn't understand the question

Keyword search returns files, not answers. If the document phrases it differently than the person asking, the right file never surfaces at all.

The corpus is scattered and inconsistent

Scanned PDFs, spreadsheets, CAD exports, email attachments, a decade of folder conventions that changed three times. No single place is authoritative.

Not everyone may see everything

Commercial terms, HR files and client-confidential material sit in the same archive as the general documentation. A search tool that ignores permissions is a liability, not a feature.

A generic chatbot will make things up

An LLM pointed at your data without grounding, evaluation and refusal behaviour produces confident answers that are wrong — which is worse than no answer in a regulated or contractual context.

Six pieces. Each one is where naive RAG fails.

The architecture below is the one running in production on the deployments referenced on this page. It is deliberately unexciting: the hard parts are ingestion quality, permissions and evaluation, not the model choice.

01

Ingestion that preserves structure

Documents are parsed with layout awareness rather than flattened to raw text, so tables, headings and figure references survive. On the German technical deployment this matters directly: DIN, EN and DWA norm references are preserved per chunk, because an answer that quotes a standard without saying which standard is useless.

02

Hybrid retrieval, then reranking

Dense embeddings catch paraphrase; sparse keyword retrieval catches part numbers, norm codes and exact identifiers that embeddings routinely miss. Results are merged and passed through a cross-encoder reranker before anything reaches the model.

03

Permissions enforced at retrieval time

Access control is applied when documents are fetched, not filtered out of the answer afterwards. The construction deployment uses fine-grained authorization behind SSO, so the same question asked by two people returns answers drawn only from what each of them is entitled to see.

04

Grounded generation with citations

The model answers only from retrieved passages and returns the source document alongside the answer. Users verify in one click instead of taking the system's word for it — which is what makes the tool adoptable in a business where being wrong has contractual consequences.

05

Explicit refusal when the answer isn't there

Responses are schema-constrained so the system can return "not found" as a valid outcome. In the Enterprise RAG Challenge benchmark this behaviour was measured directly: on the 18% of questions where the document contained no answer, the system refused rather than inventing one.

06

A golden evaluation set, from day one

Every deployment gets a curated set of real questions with known-correct answers — 23+ golden eval cases per tenant on the German multi-tenant system. Nothing ships unless the eval suite passes, and the suite is what tells you whether a prompt or model change actually helped.

A system you own, not a demo.

Delivery is scoped to your corpus and your access model. What's constant is the shape of what's handed over.

An indexed, permissioned corpus

Your documents ingested, chunked and indexed, with access control mapped to your existing identity provider so the system inherits the permissions you already maintain.

A search interface your team will actually use

Natural-language questions, answers with citations, one-click access to the source document. Deployed as its own app or embedded where people already work.

An evaluation suite

A golden set of questions and expected answers, plus the harness to run it. This is how you verify quality before rollout and how you catch regressions after.

Re-indexing that keeps up

New and changed documents flow into the index without a rebuild, so the answers reflect the current state of the archive rather than the day you launched.

Deployment on your terms

Cloud, your own VPC, or an on-premise setup where the corpus cannot leave your network. The retrieval layer is model-agnostic, so the LLM choice stays yours.

Handover and documentation

Architecture, ingestion pipeline and eval process documented so your engineers can extend it. No dependency on us to add a document source.

The questions that come up first.

Will it leak documents to people who shouldn't see them?

Not if permissions are enforced at retrieval. That is the design: documents a user isn't entitled to are never fetched, so they can't appear in an answer or a citation. The construction deployment runs fine-grained authorization behind SSO for roughly 500 internal users.

Does our data go to OpenAI or Anthropic?

Only if you want it to. The retrieval layer is model-agnostic — deployments have run on managed cloud models and can run against models hosted inside your own network. The corpus and the index live where you decide.

What about hallucinations?

Two mechanisms. Answers are generated only from retrieved passages and shipped with citations, so a wrong answer is visible rather than plausible. And responses are schema-constrained so "the documents don't contain this" is a valid output — measured at 0 hallucinations in the Enterprise RAG Challenge, where the system refused on the 18% of questions that had no answer in the source.

Our documents are scanned PDFs and old CAD exports. Does that work?

It's the normal case, and it's where most of the effort goes. Layout-aware parsing handles tables and structured pages; scanned material goes through OCR before chunking. The honest answer is that ingestion quality determines answer quality, so we scope it against a sample of your actual worst files before committing.

How do we know it's good enough to roll out?

The golden evaluation set. Real questions from your team with known-correct answers, run as a suite. You see the pass rate before rollout rather than discovering quality problems from users. On the German multi-tenant system each tenant carries 23+ such cases.

What does it cost?

It depends on the corpus, the number of source systems and the access model — those drive the work far more than the interface does. We scope it on a call after looking at a representative sample of your documents. If you want a published price range for the customer-facing variant, the AI chatbots service page lists it.

Bring a sample of your worst documents.

A 30-minute call. We look at what you actually have — the scanned specs, the inconsistent folders, the permission model — and tell you whether retrieval will work well on it and what it would take. No slides.

Book a call →

Response within 24h · Bucharest, Romania