Crescentek

AI Knowledge Base & Search

Ask your company brain.

Your SOPs, contracts, playbooks, and project archives — searchable by meaning, not just keywords. Every answer cites the source document, page, and section. Built on RAG (retrieval-augmented generation), private to your org.

Cites every source
Private to your org
EU-hosted vectors
Searching 2,847 documents…
What you connect

The places knowledge actually lives.

Most companies don't lack documentation. They lack the ability to find it across 6 different tools. We unify the search layer without moving the documents.

SharePoint / OneDrive
Google Drive
Notion / Confluence
Local PDFs / DOCX
CRM notes (HubSpot, Salesforce)
Slack / Teams threads
Internal wikis / Intranets
Helpdesk articles
Under the hood

Retrieval-Augmented Generation, plainly.

An LLM on its own knows nothing about your company — and inventing an answer is the failure mode you must avoid. RAG fetches the relevant chunks of your documents first, then asks the LLM to answer using only those. Citations are the proof of work.

01
Ingest
Read all your documents from each source. Connectors keep them in sync as they change.
02
Chunk + embed
Each document split into semantic chunks. Each chunk vectorised so similar meanings cluster together.
03
Query
User asks a question. We embed the question the same way, and retrieve the closest-meaning chunks.
04
Ground
LLM receives only the retrieved chunks + the question. Must answer from those — refusing if there's no support.
05
Cite
Every claim links back to its source doc + page. Verifiable, auditable, GDPR-friendly.
Built on

The RAG stack we deploy.

PineconeWeaviatepgvector (Postgres)QdrantClaude (Anthropic)OpenAI GPT-4oOpenAI embeddingsVoyage embeddingsLangChainLlamaIndexSharePoint Graph APIGoogle Drive APINotion API
Where RAG falls flat

Three things to fix before deploying.

Stale or conflicting docs
If you have 5 versions of the same policy in different folders, RAG will retrieve all 5 and confuse the answer. We help identify the source of truth before indexing. Often the deliverable is documentation cleanup as much as the AI build.
Permission boundaries
Junior staff shouldn't get answers from senior-only documents. We mirror your existing access controls (SharePoint permissions, Drive sharing rules) at query time — answers respect who's asking.
Sensitive data exposure
If your docs contain PII, the search experience can leak. We support per-document redaction, sensitive-content classifiers, and field-level access rules. Required for healthcare and finance use cases.
Frequently asked

Knowledge base questions.

Native search matches keywords in filenames and content. RAG matches meaning — 'how do refunds work for B2B?' returns the right contract clause even if it never uses the word 'refund'. And the answer is synthesised from multiple sources with citations, not just a list of links.
Vectors in your chosen DB (Pinecone Ireland, pgvector on your existing Postgres, Qdrant self-hosted — your call). Original documents stay where they are (SharePoint, Drive, etc.) — we don't copy them to a separate store. EU-hosted by default.
Webhook-based for tools that support it (Google Drive, Notion). Scheduled re-indexing (every 2-15 minutes) for those that don't. Stale answers are the worst failure mode, so freshness is a first-class concern.
Yes — and we configure it to refuse to answer if no sufficient source is found, rather than hallucinate. 'I couldn't find a confident answer in the knowledge base' is a feature, not a failure.
Yes — at query time. If User A doesn't have access to a document, the chunks from that document aren't retrieved for their query. Built into the retrieval layer, not bolted on.
Build: €6,000–€20,000 depending on source count and complexity. Ongoing: €100–€500/mo (embeddings + LLM + vector hosting). We price per document volume during scoping for predictability.

How many times this week did someone ask: "where's that doc?"

That's the problem we solve. Drop us 50 of your real documents and we'll show you what a private RAG search looks like over your content.