AI · RAG

Knowledge chatbots that don't lie.

RAG chatbots over your real documentation, policies, or product knowledge. Cited, monitored, evaluated — not a wrapped GPT-4 demo. We build the kind that survives the post-launch scrutiny.

And why most RAG demos fail in production.

What RAG actually is

RAG (Retrieval-Augmented Generation) is a pattern where the LLM doesn't answer from training data — it answers from documents you control. The system retrieves relevant chunks of your knowledge base, then asks the model to answer based only on those chunks, with citations.

Built right, RAG dramatically reduces hallucinations and gives users a reason to trust the output. Built sloppily, it produces plausible-sounding answers that miss the actual document, cite the wrong source, or confidently make things up when the answer isn't in the knowledge base.

The difference is engineering: chunking strategy, embeddings choice, retrieval evaluation, prompt design, fallback behaviour, and continuous monitoring. We treat RAG as a real software system, not a weekend project.

Where RAG genuinely earns its place.

What we build

Internal knowledge chatbots

Over policies, SOPs, product docs, sales playbooks. Help your team find answers without pinging Slack — with citations they can verify.

Customer-facing support agents

Answers grounded in your real documentation. Knows when to escalate. Logs every conversation for review.

Sales enablement assistants

Sales-specific knowledge bots that pull from product docs, competitive intel, and past deals. Cite their sources to your sales team.

Compliance & policy assistants

RAG over regulations, internal policies, contract clauses. Useful for legal, HR, finance teams. Citations are non-negotiable.

Multi-source agentic RAG

Bots that decide which knowledge base to consult, ask clarifying questions, and combine retrieval with light agency where it makes sense.

Six things that separate a real system from a demo.

How we build RAG that works

01

Curate the knowledge base

Garbage in = garbage out. We help you decide what's in scope, what's stale, and what needs to be rewritten before a chatbot ever sees it.

02

Chunking & retrieval that fits your content

Different content types need different chunking. Legal docs ≠ FAQs ≠ wikis. We tune retrieval per knowledge base, not with defaults.

03

Citations and 'I don't know'

Every answer cites sources. The bot says 'I don't know' or escalates when the answer isn't in the corpus — instead of inventing.

04

Evaluation: golden test sets

We build a test set of representative questions with expected sources. Every model or prompt change is scored. No vibes-driven RAG.

05

Monitoring in production

All conversations logged. Drift, refusal rate, citation accuracy tracked. You see what users actually ask versus what you thought they would.

What a RAG chatbot costs

Pricing

€20–80K

scoped per knowledge base · 4–8 weeks

A production RAG chatbot over a defined knowledge base typically costs €20–50K including evaluation framework and monitoring. More complex setups — multiple knowledge bases, role-based access, agentic routing — €50–80K. Plus model and infrastructure costs (€50–500/mo depending on volume).

RAG chatbots running in production today.

What we've shipped

REAL ESTATE

Tenant FAQ bot

Answers tenant questions over the leasing handbook, building rules, and FAQs. Cuts inbound tenant queries to property managers by ~40%.

B2B SERVICES

Sales enablement assistant

Sales team chatbot over product docs, competitive intel, and recent customer wins. Cited responses; logs feedback to improve sources over time.

INTERNAL TOOLS

HR policy chatbot

Answers employee questions about leave, benefits, expenses. Always cites the policy. Escalates ambiguous questions to HR.

Common questions

A RAG chatbot answers questions over a knowledge base. An AI agent takes actions (calls APIs, writes data, executes workflows). Many systems combine both.

Sometimes. If your knowledge is broadly available on the public internet, general models often suffice. RAG earns its keep when (a) your knowledge is private, (b) you need traceable citations, (c) you need control over what's said, or (d) you want measurement and drift detection — none of which generic chatbots provide.

We've built systems over knowledge bases from a few hundred to tens of thousands of documents. Architecture changes with scale (vector DB choice, retrieval strategy, indexing pipeline) but neither extreme is unusual.

Sort of. We log every interaction and surface patterns where the bot got it wrong. Most fixes are improvements to the knowledge base itself or to retrieval — not 'training' the model. Continuous improvement is operational, not magical.

Per knowledge base: answer accuracy on the eval set, citation accuracy, refusal rate (saying 'I don't know' when the answer isn't there), user satisfaction, and downstream impact (e.g., reduction in tickets to humans).

Pair with

Have a knowledge base your team needs to query faster?

Tell us what knowledge and who'd use it. We'll come back within one business day with whether RAG fits and where to start.